id
stringlengths
11
11
created
timestamp[s]date
2026-01-01 00:00:00
2026-01-01 00:00:00
topic
stringclasses
12 values
task_type
stringclasses
8 values
difficulty
stringclasses
4 values
instruction
stringlengths
201
264
input
stringclasses
1 value
output
stringclasses
7 values
metadata
dict
train_00300
2026-01-01T00:00:00
Dataset curation pipelines (filter, dedupe, quality)
code
advanced
Task: code Topic: Dataset curation pipelines (filter, dedupe, quality) Difficulty: advanced Target language: C# Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "C#", "developer_needs": [ "governance", "reproducibility", "cost_latency_tradeoffs" ] }
train_00301
2026-01-01T00:00:00
Model merging, distillation, and continued pretraining
data_pipeline
advanced
Task: data_pipeline Topic: Model merging, distillation, and continued pretraining Difficulty: advanced Target language: Rust Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Pipeline: 1) Ingest 2) Normalize 3) Filter 4) Dedupe 5) Quality score 6) Sample 7) Audit
{ "target_language": "Rust", "developer_needs": [ "reproducibility", "governance", "tooling" ] }
train_00302
2026-01-01T00:00:00
Governance, provenance, and licensing for code data
explain
advanced
Task: explain Topic: Governance, provenance, and licensing for code data Difficulty: advanced Target language: C# Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "C#", "developer_needs": [ "repo_scale_reasoning", "ci_integration", "evaluation_metrics" ] }
train_00303
2026-01-01T00:00:00
Reasoning-first coding models and tunable deliberation
compare
expert
Task: compare Topic: Reasoning-first coding models and tunable deliberation Difficulty: expert Target language: Bash Context: Design a data pipeline for continued pretraining with auditability. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Compare: capability, cost, latency, reliability, governance
{ "target_language": "Bash", "developer_needs": [ "tooling", "governance", "repo_scale_reasoning" ] }
train_00304
2026-01-01T00:00:00
Multimodal dev workflows (docs, diagrams, traces)
design
intermediate
Task: design Topic: Multimodal dev workflows (docs, diagrams, traces) Difficulty: intermediate Target language: Go Context: Design a data pipeline for continued pretraining with auditability. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Go", "developer_needs": [ "governance", "ci_integration", "tests_are_truth" ] }
train_00305
2026-01-01T00:00:00
Multimodal dev workflows (docs, diagrams, traces)
explain
foundation
Task: explain Topic: Multimodal dev workflows (docs, diagrams, traces) Difficulty: foundation Target language: Bash Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Bash", "developer_needs": [ "documentation", "ci_integration", "tests_are_truth" ] }
train_00306
2026-01-01T00:00:00
Agentic coding systems (plan→edit→test→reflect)
explain
foundation
Task: explain Topic: Agentic coding systems (plan→edit→test→reflect) Difficulty: foundation Target language: Python Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Python", "developer_needs": [ "security_gates", "evaluation_metrics", "repo_scale_reasoning" ] }
train_00307
2026-01-01T00:00:00
Secure code generation and policy gates
eval
foundation
Task: eval Topic: Secure code generation and policy gates Difficulty: foundation Target language: JavaScript Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "JavaScript", "developer_needs": [ "tests_are_truth", "evaluation_metrics", "governance" ] }
train_00308
2026-01-01T00:00:00
Mixture-of-Experts (MoE) for code
eval
foundation
Task: eval Topic: Mixture-of-Experts (MoE) for code Difficulty: foundation Target language: SQL Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "SQL", "developer_needs": [ "documentation", "tests_are_truth", "ci_integration" ] }
train_00309
2026-01-01T00:00:00
Governance, provenance, and licensing for code data
agent_loop
advanced
Task: agent_loop Topic: Governance, provenance, and licensing for code data Difficulty: advanced Target language: C# Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Agent loop: Plan → Edit (diff) → Test → Reflect → Human gate
{ "target_language": "C#", "developer_needs": [ "documentation", "cost_latency_tradeoffs", "governance" ] }
train_00310
2026-01-01T00:00:00
Extended context and repo-scale understanding
data_pipeline
foundation
Task: data_pipeline Topic: Extended context and repo-scale understanding Difficulty: foundation Target language: TypeScript Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Pipeline: 1) Ingest 2) Normalize 3) Filter 4) Dedupe 5) Quality score 6) Sample 7) Audit
{ "target_language": "TypeScript", "developer_needs": [ "governance", "tooling", "evaluation_metrics" ] }
train_00311
2026-01-01T00:00:00
Tool calling, sandboxes, and CI integration
design
intermediate
Task: design Topic: Tool calling, sandboxes, and CI integration Difficulty: intermediate Target language: Python Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Python", "developer_needs": [ "tooling", "ci_integration", "repo_scale_reasoning" ] }
train_00312
2026-01-01T00:00:00
Model merging, distillation, and continued pretraining
data_pipeline
foundation
Task: data_pipeline Topic: Model merging, distillation, and continued pretraining Difficulty: foundation Target language: Go Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Pipeline: 1) Ingest 2) Normalize 3) Filter 4) Dedupe 5) Quality score 6) Sample 7) Audit
{ "target_language": "Go", "developer_needs": [ "reproducibility", "security_gates", "documentation" ] }
train_00313
2026-01-01T00:00:00
Mixture-of-Experts (MoE) for code
eval
foundation
Task: eval Topic: Mixture-of-Experts (MoE) for code Difficulty: foundation Target language: C# Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "C#", "developer_needs": [ "ci_integration", "governance", "security_gates" ] }
train_00314
2026-01-01T00:00:00
Multimodal dev workflows (docs, diagrams, traces)
review
expert
Task: review Topic: Multimodal dev workflows (docs, diagrams, traces) Difficulty: expert Target language: JavaScript Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Review: correctness, security, performance, governance
{ "target_language": "JavaScript", "developer_needs": [ "tooling", "repo_scale_reasoning", "cost_latency_tradeoffs" ] }
train_00315
2026-01-01T00:00:00
Secure code generation and policy gates
design
intermediate
Task: design Topic: Secure code generation and policy gates Difficulty: intermediate Target language: Bash Context: Design a data pipeline for continued pretraining with auditability. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Bash", "developer_needs": [ "evaluation_metrics", "repo_scale_reasoning", "reproducibility" ] }
train_00316
2026-01-01T00:00:00
Code-specialized model families and sizing tradeoffs
eval
foundation
Task: eval Topic: Code-specialized model families and sizing tradeoffs Difficulty: foundation Target language: Rust Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "Rust", "developer_needs": [ "security_gates", "governance", "documentation" ] }
train_00317
2026-01-01T00:00:00
Reasoning-first coding models and tunable deliberation
data_pipeline
foundation
Task: data_pipeline Topic: Reasoning-first coding models and tunable deliberation Difficulty: foundation Target language: Python Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Pipeline: 1) Ingest 2) Normalize 3) Filter 4) Dedupe 5) Quality score 6) Sample 7) Audit
{ "target_language": "Python", "developer_needs": [ "evaluation_metrics", "repo_scale_reasoning", "cost_latency_tradeoffs" ] }
train_00318
2026-01-01T00:00:00
Secure code generation and policy gates
data_pipeline
foundation
Task: data_pipeline Topic: Secure code generation and policy gates Difficulty: foundation Target language: Bash Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Pipeline: 1) Ingest 2) Normalize 3) Filter 4) Dedupe 5) Quality score 6) Sample 7) Audit
{ "target_language": "Bash", "developer_needs": [ "evaluation_metrics", "tests_are_truth", "governance" ] }
train_00319
2026-01-01T00:00:00
Code-specialized model families and sizing tradeoffs
eval
foundation
Task: eval Topic: Code-specialized model families and sizing tradeoffs Difficulty: foundation Target language: JavaScript Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "JavaScript", "developer_needs": [ "security_gates", "tooling", "documentation" ] }
train_00320
2026-01-01T00:00:00
Multimodal dev workflows (docs, diagrams, traces)
code
advanced
Task: code Topic: Multimodal dev workflows (docs, diagrams, traces) Difficulty: advanced Target language: TypeScript Context: Design a data pipeline for continued pretraining with auditability. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "TypeScript", "developer_needs": [ "tests_are_truth", "reproducibility", "repo_scale_reasoning" ] }
train_00321
2026-01-01T00:00:00
SWE-bench style real-repo evaluation
code
foundation
Task: code Topic: SWE-bench style real-repo evaluation Difficulty: foundation Target language: Go Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Go", "developer_needs": [ "evaluation_metrics", "tests_are_truth", "repo_scale_reasoning" ] }
train_00322
2026-01-01T00:00:00
SWE-bench style real-repo evaluation
review
expert
Task: review Topic: SWE-bench style real-repo evaluation Difficulty: expert Target language: C# Context: Design a data pipeline for continued pretraining with auditability. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Review: correctness, security, performance, governance
{ "target_language": "C#", "developer_needs": [ "governance", "cost_latency_tradeoffs", "documentation" ] }
train_00323
2026-01-01T00:00:00
Reasoning-first coding models and tunable deliberation
eval
intermediate
Task: eval Topic: Reasoning-first coding models and tunable deliberation Difficulty: intermediate Target language: Go Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "Go", "developer_needs": [ "reproducibility", "security_gates", "cost_latency_tradeoffs" ] }
train_00324
2026-01-01T00:00:00
Reasoning-first coding models and tunable deliberation
data_pipeline
foundation
Task: data_pipeline Topic: Reasoning-first coding models and tunable deliberation Difficulty: foundation Target language: Bash Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Pipeline: 1) Ingest 2) Normalize 3) Filter 4) Dedupe 5) Quality score 6) Sample 7) Audit
{ "target_language": "Bash", "developer_needs": [ "tests_are_truth", "documentation", "evaluation_metrics" ] }
train_00325
2026-01-01T00:00:00
Reasoning-first coding models and tunable deliberation
eval
expert
Task: eval Topic: Reasoning-first coding models and tunable deliberation Difficulty: expert Target language: C# Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "C#", "developer_needs": [ "tests_are_truth", "evaluation_metrics", "tooling" ] }
train_00326
2026-01-01T00:00:00
Multimodal dev workflows (docs, diagrams, traces)
review
foundation
Task: review Topic: Multimodal dev workflows (docs, diagrams, traces) Difficulty: foundation Target language: Python Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Review: correctness, security, performance, governance
{ "target_language": "Python", "developer_needs": [ "documentation", "reproducibility", "tests_are_truth" ] }
train_00327
2026-01-01T00:00:00
SWE-bench style real-repo evaluation
data_pipeline
advanced
Task: data_pipeline Topic: SWE-bench style real-repo evaluation Difficulty: advanced Target language: TypeScript Context: Design a data pipeline for continued pretraining with auditability. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Pipeline: 1) Ingest 2) Normalize 3) Filter 4) Dedupe 5) Quality score 6) Sample 7) Audit
{ "target_language": "TypeScript", "developer_needs": [ "governance", "tooling", "evaluation_metrics" ] }
train_00328
2026-01-01T00:00:00
Code-specialized model families and sizing tradeoffs
eval
intermediate
Task: eval Topic: Code-specialized model families and sizing tradeoffs Difficulty: intermediate Target language: Java Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "Java", "developer_needs": [ "cost_latency_tradeoffs", "tests_are_truth", "reproducibility" ] }
train_00329
2026-01-01T00:00:00
Reasoning-first coding models and tunable deliberation
eval
intermediate
Task: eval Topic: Reasoning-first coding models and tunable deliberation Difficulty: intermediate Target language: JavaScript Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "JavaScript", "developer_needs": [ "reproducibility", "evaluation_metrics", "ci_integration" ] }
train_00330
2026-01-01T00:00:00
Reasoning-first coding models and tunable deliberation
code
advanced
Task: code Topic: Reasoning-first coding models and tunable deliberation Difficulty: advanced Target language: Bash Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Bash", "developer_needs": [ "tooling", "repo_scale_reasoning", "governance" ] }
train_00331
2026-01-01T00:00:00
Dataset curation pipelines (filter, dedupe, quality)
explain
expert
Task: explain Topic: Dataset curation pipelines (filter, dedupe, quality) Difficulty: expert Target language: TypeScript Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "TypeScript", "developer_needs": [ "reproducibility", "cost_latency_tradeoffs", "ci_integration" ] }
train_00332
2026-01-01T00:00:00
Mixture-of-Experts (MoE) for code
data_pipeline
intermediate
Task: data_pipeline Topic: Mixture-of-Experts (MoE) for code Difficulty: intermediate Target language: C# Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Pipeline: 1) Ingest 2) Normalize 3) Filter 4) Dedupe 5) Quality score 6) Sample 7) Audit
{ "target_language": "C#", "developer_needs": [ "cost_latency_tradeoffs", "evaluation_metrics", "reproducibility" ] }
train_00333
2026-01-01T00:00:00
SWE-bench style real-repo evaluation
compare
expert
Task: compare Topic: SWE-bench style real-repo evaluation Difficulty: expert Target language: TypeScript Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Compare: capability, cost, latency, reliability, governance
{ "target_language": "TypeScript", "developer_needs": [ "security_gates", "tooling", "documentation" ] }
train_00334
2026-01-01T00:00:00
SWE-bench style real-repo evaluation
review
intermediate
Task: review Topic: SWE-bench style real-repo evaluation Difficulty: intermediate Target language: TypeScript Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Review: correctness, security, performance, governance
{ "target_language": "TypeScript", "developer_needs": [ "ci_integration", "repo_scale_reasoning", "tooling" ] }
train_00335
2026-01-01T00:00:00
Extended context and repo-scale understanding
explain
foundation
Task: explain Topic: Extended context and repo-scale understanding Difficulty: foundation Target language: Go Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Go", "developer_needs": [ "governance", "tests_are_truth", "tooling" ] }
train_00336
2026-01-01T00:00:00
Model merging, distillation, and continued pretraining
data_pipeline
intermediate
Task: data_pipeline Topic: Model merging, distillation, and continued pretraining Difficulty: intermediate Target language: Java Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Pipeline: 1) Ingest 2) Normalize 3) Filter 4) Dedupe 5) Quality score 6) Sample 7) Audit
{ "target_language": "Java", "developer_needs": [ "cost_latency_tradeoffs", "ci_integration", "tooling" ] }
train_00337
2026-01-01T00:00:00
Extended context and repo-scale understanding
design
advanced
Task: design Topic: Extended context and repo-scale understanding Difficulty: advanced Target language: JavaScript Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "JavaScript", "developer_needs": [ "reproducibility", "ci_integration", "tests_are_truth" ] }
train_00338
2026-01-01T00:00:00
Dataset curation pipelines (filter, dedupe, quality)
code
advanced
Task: code Topic: Dataset curation pipelines (filter, dedupe, quality) Difficulty: advanced Target language: Python Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Reference scaffold: ```python def agent_loop(plan, edit, test, issue, max_iters=3): history = [] p = plan(issue) for _ in range(max_iters): patch = edit(issue, p) ok, report = test(patch) history.append({"plan": p, "passed": ok, "report": report[:200]}) if ok: return patch, history p = p + " | refine from failures" return patch, history ``` Operational notes: sandbox, pinned deps, human gate.
{ "target_language": "Python", "developer_needs": [ "ci_integration", "reproducibility", "security_gates" ] }
train_00339
2026-01-01T00:00:00
Governance, provenance, and licensing for code data
agent_loop
foundation
Task: agent_loop Topic: Governance, provenance, and licensing for code data Difficulty: foundation Target language: Rust Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Agent loop: Plan → Edit (diff) → Test → Reflect → Human gate
{ "target_language": "Rust", "developer_needs": [ "reproducibility", "security_gates", "evaluation_metrics" ] }
train_00340
2026-01-01T00:00:00
Multimodal dev workflows (docs, diagrams, traces)
review
intermediate
Task: review Topic: Multimodal dev workflows (docs, diagrams, traces) Difficulty: intermediate Target language: TypeScript Context: Design a data pipeline for continued pretraining with auditability. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Review: correctness, security, performance, governance
{ "target_language": "TypeScript", "developer_needs": [ "governance", "security_gates", "documentation" ] }
train_00341
2026-01-01T00:00:00
Secure code generation and policy gates
code
advanced
Task: code Topic: Secure code generation and policy gates Difficulty: advanced Target language: Python Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Reference scaffold: ```python def agent_loop(plan, edit, test, issue, max_iters=3): history = [] p = plan(issue) for _ in range(max_iters): patch = edit(issue, p) ok, report = test(patch) history.append({"plan": p, "passed": ok, "report": report[:200]}) if ok: return patch, history p = p + " | refine from failures" return patch, history ``` Operational notes: sandbox, pinned deps, human gate.
{ "target_language": "Python", "developer_needs": [ "tests_are_truth", "cost_latency_tradeoffs", "evaluation_metrics" ] }
train_00342
2026-01-01T00:00:00
Tool calling, sandboxes, and CI integration
data_pipeline
foundation
Task: data_pipeline Topic: Tool calling, sandboxes, and CI integration Difficulty: foundation Target language: Python Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Pipeline: 1) Ingest 2) Normalize 3) Filter 4) Dedupe 5) Quality score 6) Sample 7) Audit
{ "target_language": "Python", "developer_needs": [ "ci_integration", "repo_scale_reasoning", "reproducibility" ] }
train_00343
2026-01-01T00:00:00
Extended context and repo-scale understanding
agent_loop
intermediate
Task: agent_loop Topic: Extended context and repo-scale understanding Difficulty: intermediate Target language: JavaScript Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Agent loop: Plan → Edit (diff) → Test → Reflect → Human gate
{ "target_language": "JavaScript", "developer_needs": [ "evaluation_metrics", "reproducibility", "cost_latency_tradeoffs" ] }
train_00344
2026-01-01T00:00:00
SWE-bench style real-repo evaluation
compare
expert
Task: compare Topic: SWE-bench style real-repo evaluation Difficulty: expert Target language: Java Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Compare: capability, cost, latency, reliability, governance
{ "target_language": "Java", "developer_needs": [ "documentation", "evaluation_metrics", "ci_integration" ] }
train_00345
2026-01-01T00:00:00
Multimodal dev workflows (docs, diagrams, traces)
compare
expert
Task: compare Topic: Multimodal dev workflows (docs, diagrams, traces) Difficulty: expert Target language: Python Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Compare: capability, cost, latency, reliability, governance
{ "target_language": "Python", "developer_needs": [ "governance", "documentation", "evaluation_metrics" ] }
train_00346
2026-01-01T00:00:00
SWE-bench style real-repo evaluation
eval
advanced
Task: eval Topic: SWE-bench style real-repo evaluation Difficulty: advanced Target language: Python Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "Python", "developer_needs": [ "tooling", "repo_scale_reasoning", "reproducibility" ] }
train_00347
2026-01-01T00:00:00
SWE-bench style real-repo evaluation
compare
advanced
Task: compare Topic: SWE-bench style real-repo evaluation Difficulty: advanced Target language: SQL Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Compare: capability, cost, latency, reliability, governance
{ "target_language": "SQL", "developer_needs": [ "tooling", "governance", "reproducibility" ] }
train_00348
2026-01-01T00:00:00
Dataset curation pipelines (filter, dedupe, quality)
compare
expert
Task: compare Topic: Dataset curation pipelines (filter, dedupe, quality) Difficulty: expert Target language: Bash Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Compare: capability, cost, latency, reliability, governance
{ "target_language": "Bash", "developer_needs": [ "evaluation_metrics", "reproducibility", "tooling" ] }
train_00349
2026-01-01T00:00:00
Agentic coding systems (plan→edit→test→reflect)
agent_loop
advanced
Task: agent_loop Topic: Agentic coding systems (plan→edit→test→reflect) Difficulty: advanced Target language: Python Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Agent loop: Plan → Edit (diff) → Test → Reflect → Human gate
{ "target_language": "Python", "developer_needs": [ "tooling", "reproducibility", "tests_are_truth" ] }
train_00350
2026-01-01T00:00:00
Governance, provenance, and licensing for code data
design
expert
Task: design Topic: Governance, provenance, and licensing for code data Difficulty: expert Target language: Go Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Go", "developer_needs": [ "reproducibility", "repo_scale_reasoning", "governance" ] }
train_00351
2026-01-01T00:00:00
Governance, provenance, and licensing for code data
data_pipeline
foundation
Task: data_pipeline Topic: Governance, provenance, and licensing for code data Difficulty: foundation Target language: Java Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Pipeline: 1) Ingest 2) Normalize 3) Filter 4) Dedupe 5) Quality score 6) Sample 7) Audit
{ "target_language": "Java", "developer_needs": [ "governance", "repo_scale_reasoning", "security_gates" ] }
train_00352
2026-01-01T00:00:00
Tool calling, sandboxes, and CI integration
eval
intermediate
Task: eval Topic: Tool calling, sandboxes, and CI integration Difficulty: intermediate Target language: JavaScript Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "JavaScript", "developer_needs": [ "reproducibility", "tests_are_truth", "governance" ] }
train_00353
2026-01-01T00:00:00
Mixture-of-Experts (MoE) for code
explain
intermediate
Task: explain Topic: Mixture-of-Experts (MoE) for code Difficulty: intermediate Target language: Bash Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Bash", "developer_needs": [ "evaluation_metrics", "tooling", "reproducibility" ] }
train_00354
2026-01-01T00:00:00
Mixture-of-Experts (MoE) for code
explain
intermediate
Task: explain Topic: Mixture-of-Experts (MoE) for code Difficulty: intermediate Target language: Bash Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Bash", "developer_needs": [ "repo_scale_reasoning", "ci_integration", "cost_latency_tradeoffs" ] }
train_00355
2026-01-01T00:00:00
Agentic coding systems (plan→edit→test→reflect)
eval
intermediate
Task: eval Topic: Agentic coding systems (plan→edit→test→reflect) Difficulty: intermediate Target language: Java Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "Java", "developer_needs": [ "tooling", "evaluation_metrics", "cost_latency_tradeoffs" ] }
train_00356
2026-01-01T00:00:00
Multimodal dev workflows (docs, diagrams, traces)
data_pipeline
foundation
Task: data_pipeline Topic: Multimodal dev workflows (docs, diagrams, traces) Difficulty: foundation Target language: Java Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Pipeline: 1) Ingest 2) Normalize 3) Filter 4) Dedupe 5) Quality score 6) Sample 7) Audit
{ "target_language": "Java", "developer_needs": [ "cost_latency_tradeoffs", "reproducibility", "ci_integration" ] }
train_00357
2026-01-01T00:00:00
Secure code generation and policy gates
review
advanced
Task: review Topic: Secure code generation and policy gates Difficulty: advanced Target language: Go Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Review: correctness, security, performance, governance
{ "target_language": "Go", "developer_needs": [ "cost_latency_tradeoffs", "repo_scale_reasoning", "governance" ] }
train_00358
2026-01-01T00:00:00
Multimodal dev workflows (docs, diagrams, traces)
compare
advanced
Task: compare Topic: Multimodal dev workflows (docs, diagrams, traces) Difficulty: advanced Target language: Python Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Compare: capability, cost, latency, reliability, governance
{ "target_language": "Python", "developer_needs": [ "tooling", "security_gates", "governance" ] }
train_00359
2026-01-01T00:00:00
Code-specialized model families and sizing tradeoffs
explain
advanced
Task: explain Topic: Code-specialized model families and sizing tradeoffs Difficulty: advanced Target language: SQL Context: Design a data pipeline for continued pretraining with auditability. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "SQL", "developer_needs": [ "evaluation_metrics", "tooling", "security_gates" ] }
train_00360
2026-01-01T00:00:00
Mixture-of-Experts (MoE) for code
agent_loop
foundation
Task: agent_loop Topic: Mixture-of-Experts (MoE) for code Difficulty: foundation Target language: TypeScript Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Agent loop: Plan → Edit (diff) → Test → Reflect → Human gate
{ "target_language": "TypeScript", "developer_needs": [ "cost_latency_tradeoffs", "reproducibility", "repo_scale_reasoning" ] }
train_00361
2026-01-01T00:00:00
SWE-bench style real-repo evaluation
code
intermediate
Task: code Topic: SWE-bench style real-repo evaluation Difficulty: intermediate Target language: SQL Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "SQL", "developer_needs": [ "documentation", "security_gates", "tests_are_truth" ] }
train_00362
2026-01-01T00:00:00
Code-specialized model families and sizing tradeoffs
review
expert
Task: review Topic: Code-specialized model families and sizing tradeoffs Difficulty: expert Target language: Go Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Review: correctness, security, performance, governance
{ "target_language": "Go", "developer_needs": [ "evaluation_metrics", "security_gates", "ci_integration" ] }
train_00363
2026-01-01T00:00:00
Reasoning-first coding models and tunable deliberation
design
intermediate
Task: design Topic: Reasoning-first coding models and tunable deliberation Difficulty: intermediate Target language: C# Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "C#", "developer_needs": [ "tests_are_truth", "documentation", "repo_scale_reasoning" ] }
train_00364
2026-01-01T00:00:00
Model merging, distillation, and continued pretraining
compare
intermediate
Task: compare Topic: Model merging, distillation, and continued pretraining Difficulty: intermediate Target language: Java Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Compare: capability, cost, latency, reliability, governance
{ "target_language": "Java", "developer_needs": [ "repo_scale_reasoning", "tooling", "governance" ] }
train_00365
2026-01-01T00:00:00
Reasoning-first coding models and tunable deliberation
code
intermediate
Task: code Topic: Reasoning-first coding models and tunable deliberation Difficulty: intermediate Target language: SQL Context: Design a data pipeline for continued pretraining with auditability. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "SQL", "developer_needs": [ "repo_scale_reasoning", "tests_are_truth", "documentation" ] }
train_00366
2026-01-01T00:00:00
Multimodal dev workflows (docs, diagrams, traces)
eval
foundation
Task: eval Topic: Multimodal dev workflows (docs, diagrams, traces) Difficulty: foundation Target language: TypeScript Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "TypeScript", "developer_needs": [ "evaluation_metrics", "documentation", "cost_latency_tradeoffs" ] }
train_00367
2026-01-01T00:00:00
SWE-bench style real-repo evaluation
code
foundation
Task: code Topic: SWE-bench style real-repo evaluation Difficulty: foundation Target language: Java Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Java", "developer_needs": [ "tests_are_truth", "documentation", "evaluation_metrics" ] }
train_00368
2026-01-01T00:00:00
Secure code generation and policy gates
review
expert
Task: review Topic: Secure code generation and policy gates Difficulty: expert Target language: Bash Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Review: correctness, security, performance, governance
{ "target_language": "Bash", "developer_needs": [ "repo_scale_reasoning", "ci_integration", "tooling" ] }
train_00369
2026-01-01T00:00:00
Tool calling, sandboxes, and CI integration
eval
advanced
Task: eval Topic: Tool calling, sandboxes, and CI integration Difficulty: advanced Target language: Python Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "Python", "developer_needs": [ "tests_are_truth", "governance", "tooling" ] }
train_00370
2026-01-01T00:00:00
Secure code generation and policy gates
agent_loop
expert
Task: agent_loop Topic: Secure code generation and policy gates Difficulty: expert Target language: Java Context: Design a data pipeline for continued pretraining with auditability. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Agent loop: Plan → Edit (diff) → Test → Reflect → Human gate
{ "target_language": "Java", "developer_needs": [ "tests_are_truth", "documentation", "security_gates" ] }
train_00371
2026-01-01T00:00:00
Code-specialized model families and sizing tradeoffs
code
expert
Task: code Topic: Code-specialized model families and sizing tradeoffs Difficulty: expert Target language: Python Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Reference scaffold: ```python def agent_loop(plan, edit, test, issue, max_iters=3): history = [] p = plan(issue) for _ in range(max_iters): patch = edit(issue, p) ok, report = test(patch) history.append({"plan": p, "passed": ok, "report": report[:200]}) if ok: return patch, history p = p + " | refine from failures" return patch, history ``` Operational notes: sandbox, pinned deps, human gate.
{ "target_language": "Python", "developer_needs": [ "tests_are_truth", "ci_integration", "reproducibility" ] }
train_00372
2026-01-01T00:00:00
Mixture-of-Experts (MoE) for code
explain
expert
Task: explain Topic: Mixture-of-Experts (MoE) for code Difficulty: expert Target language: JavaScript Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "JavaScript", "developer_needs": [ "documentation", "governance", "evaluation_metrics" ] }
train_00373
2026-01-01T00:00:00
Reasoning-first coding models and tunable deliberation
agent_loop
expert
Task: agent_loop Topic: Reasoning-first coding models and tunable deliberation Difficulty: expert Target language: TypeScript Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Agent loop: Plan → Edit (diff) → Test → Reflect → Human gate
{ "target_language": "TypeScript", "developer_needs": [ "ci_integration", "tests_are_truth", "reproducibility" ] }
train_00374
2026-01-01T00:00:00
Dataset curation pipelines (filter, dedupe, quality)
agent_loop
expert
Task: agent_loop Topic: Dataset curation pipelines (filter, dedupe, quality) Difficulty: expert Target language: Python Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Agent loop: Plan → Edit (diff) → Test → Reflect → Human gate
{ "target_language": "Python", "developer_needs": [ "documentation", "ci_integration", "tooling" ] }
train_00375
2026-01-01T00:00:00
Model merging, distillation, and continued pretraining
compare
advanced
Task: compare Topic: Model merging, distillation, and continued pretraining Difficulty: advanced Target language: Rust Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Compare: capability, cost, latency, reliability, governance
{ "target_language": "Rust", "developer_needs": [ "cost_latency_tradeoffs", "repo_scale_reasoning", "reproducibility" ] }
train_00376
2026-01-01T00:00:00
Extended context and repo-scale understanding
eval
foundation
Task: eval Topic: Extended context and repo-scale understanding Difficulty: foundation Target language: Rust Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "Rust", "developer_needs": [ "ci_integration", "tooling", "repo_scale_reasoning" ] }
train_00377
2026-01-01T00:00:00
Tool calling, sandboxes, and CI integration
compare
advanced
Task: compare Topic: Tool calling, sandboxes, and CI integration Difficulty: advanced Target language: Java Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Compare: capability, cost, latency, reliability, governance
{ "target_language": "Java", "developer_needs": [ "reproducibility", "governance", "tooling" ] }
train_00378
2026-01-01T00:00:00
Extended context and repo-scale understanding
compare
expert
Task: compare Topic: Extended context and repo-scale understanding Difficulty: expert Target language: TypeScript Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Compare: capability, cost, latency, reliability, governance
{ "target_language": "TypeScript", "developer_needs": [ "cost_latency_tradeoffs", "reproducibility", "ci_integration" ] }
train_00379
2026-01-01T00:00:00
Multimodal dev workflows (docs, diagrams, traces)
explain
foundation
Task: explain Topic: Multimodal dev workflows (docs, diagrams, traces) Difficulty: foundation Target language: Python Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Python", "developer_needs": [ "security_gates", "governance", "documentation" ] }
train_00380
2026-01-01T00:00:00
Governance, provenance, and licensing for code data
explain
intermediate
Task: explain Topic: Governance, provenance, and licensing for code data Difficulty: intermediate Target language: C# Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "C#", "developer_needs": [ "tooling", "reproducibility", "governance" ] }
train_00381
2026-01-01T00:00:00
Extended context and repo-scale understanding
explain
expert
Task: explain Topic: Extended context and repo-scale understanding Difficulty: expert Target language: TypeScript Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "TypeScript", "developer_needs": [ "reproducibility", "repo_scale_reasoning", "ci_integration" ] }
train_00382
2026-01-01T00:00:00
Dataset curation pipelines (filter, dedupe, quality)
code
foundation
Task: code Topic: Dataset curation pipelines (filter, dedupe, quality) Difficulty: foundation Target language: SQL Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "SQL", "developer_needs": [ "documentation", "tests_are_truth", "security_gates" ] }
train_00383
2026-01-01T00:00:00
Code-specialized model families and sizing tradeoffs
design
intermediate
Task: design Topic: Code-specialized model families and sizing tradeoffs Difficulty: intermediate Target language: SQL Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "SQL", "developer_needs": [ "security_gates", "cost_latency_tradeoffs", "documentation" ] }
train_00384
2026-01-01T00:00:00
Dataset curation pipelines (filter, dedupe, quality)
data_pipeline
intermediate
Task: data_pipeline Topic: Dataset curation pipelines (filter, dedupe, quality) Difficulty: intermediate Target language: C# Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Pipeline: 1) Ingest 2) Normalize 3) Filter 4) Dedupe 5) Quality score 6) Sample 7) Audit
{ "target_language": "C#", "developer_needs": [ "ci_integration", "security_gates", "tests_are_truth" ] }
train_00385
2026-01-01T00:00:00
Governance, provenance, and licensing for code data
agent_loop
expert
Task: agent_loop Topic: Governance, provenance, and licensing for code data Difficulty: expert Target language: Java Context: Design a data pipeline for continued pretraining with auditability. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Agent loop: Plan → Edit (diff) → Test → Reflect → Human gate
{ "target_language": "Java", "developer_needs": [ "tests_are_truth", "reproducibility", "repo_scale_reasoning" ] }
train_00386
2026-01-01T00:00:00
Agentic coding systems (plan→edit→test→reflect)
explain
foundation
Task: explain Topic: Agentic coding systems (plan→edit→test→reflect) Difficulty: foundation Target language: Bash Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Bash", "developer_needs": [ "documentation", "tooling", "cost_latency_tradeoffs" ] }
train_00387
2026-01-01T00:00:00
Dataset curation pipelines (filter, dedupe, quality)
agent_loop
advanced
Task: agent_loop Topic: Dataset curation pipelines (filter, dedupe, quality) Difficulty: advanced Target language: TypeScript Context: Design a data pipeline for continued pretraining with auditability. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Agent loop: Plan → Edit (diff) → Test → Reflect → Human gate
{ "target_language": "TypeScript", "developer_needs": [ "documentation", "cost_latency_tradeoffs", "governance" ] }
train_00388
2026-01-01T00:00:00
Reasoning-first coding models and tunable deliberation
data_pipeline
advanced
Task: data_pipeline Topic: Reasoning-first coding models and tunable deliberation Difficulty: advanced Target language: Python Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Pipeline: 1) Ingest 2) Normalize 3) Filter 4) Dedupe 5) Quality score 6) Sample 7) Audit
{ "target_language": "Python", "developer_needs": [ "reproducibility", "evaluation_metrics", "repo_scale_reasoning" ] }
train_00389
2026-01-01T00:00:00
Model merging, distillation, and continued pretraining
eval
foundation
Task: eval Topic: Model merging, distillation, and continued pretraining Difficulty: foundation Target language: Rust Context: Design a data pipeline for continued pretraining with auditability. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "Rust", "developer_needs": [ "security_gates", "cost_latency_tradeoffs", "tests_are_truth" ] }
train_00390
2026-01-01T00:00:00
Mixture-of-Experts (MoE) for code
eval
intermediate
Task: eval Topic: Mixture-of-Experts (MoE) for code Difficulty: intermediate Target language: Rust Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "Rust", "developer_needs": [ "tests_are_truth", "cost_latency_tradeoffs", "tooling" ] }
train_00391
2026-01-01T00:00:00
Agentic coding systems (plan→edit→test→reflect)
agent_loop
intermediate
Task: agent_loop Topic: Agentic coding systems (plan→edit→test→reflect) Difficulty: intermediate Target language: JavaScript Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Agent loop: Plan → Edit (diff) → Test → Reflect → Human gate
{ "target_language": "JavaScript", "developer_needs": [ "repo_scale_reasoning", "tests_are_truth", "governance" ] }
train_00392
2026-01-01T00:00:00
Model merging, distillation, and continued pretraining
eval
foundation
Task: eval Topic: Model merging, distillation, and continued pretraining Difficulty: foundation Target language: Rust Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "Rust", "developer_needs": [ "governance", "repo_scale_reasoning", "security_gates" ] }
train_00393
2026-01-01T00:00:00
Multimodal dev workflows (docs, diagrams, traces)
eval
expert
Task: eval Topic: Multimodal dev workflows (docs, diagrams, traces) Difficulty: expert Target language: TypeScript Context: Evaluate two coding models for internal rollout under strict governance. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "TypeScript", "developer_needs": [ "cost_latency_tradeoffs", "evaluation_metrics", "security_gates" ] }
train_00394
2026-01-01T00:00:00
Tool calling, sandboxes, and CI integration
design
intermediate
Task: design Topic: Tool calling, sandboxes, and CI integration Difficulty: intermediate Target language: Bash Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Bash", "developer_needs": [ "tests_are_truth", "documentation", "ci_integration" ] }
train_00395
2026-01-01T00:00:00
Code-specialized model families and sizing tradeoffs
design
expert
Task: design Topic: Code-specialized model families and sizing tradeoffs Difficulty: expert Target language: SQL Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "SQL", "developer_needs": [ "reproducibility", "tests_are_truth", "ci_integration" ] }
train_00396
2026-01-01T00:00:00
Mixture-of-Experts (MoE) for code
eval
foundation
Task: eval Topic: Mixture-of-Experts (MoE) for code Difficulty: foundation Target language: Java Context: Create an eval harness that reflects real developer workflows. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Eval: - Tasks: real issues - Metrics: pass@k, time-to-green - Gates: lint/security
{ "target_language": "Java", "developer_needs": [ "tests_are_truth", "documentation", "evaluation_metrics" ] }
train_00397
2026-01-01T00:00:00
Secure code generation and policy gates
explain
expert
Task: explain Topic: Secure code generation and policy gates Difficulty: expert Target language: C# Context: Fix a failing issue with tests as the oracle and produce a safe patch. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "C#", "developer_needs": [ "documentation", "tooling", "governance" ] }
train_00398
2026-01-01T00:00:00
Mixture-of-Experts (MoE) for code
compare
intermediate
Task: compare Topic: Mixture-of-Experts (MoE) for code Difficulty: intermediate Target language: SQL Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Compare: capability, cost, latency, reliability, governance
{ "target_language": "SQL", "developer_needs": [ "tests_are_truth", "reproducibility", "security_gates" ] }
train_00399
2026-01-01T00:00:00
Mixture-of-Experts (MoE) for code
code
foundation
Task: code Topic: Mixture-of-Experts (MoE) for code Difficulty: foundation Target language: Rust Context: Integrate an LLM agent into CI for a large monorepo. Deliver production-grade guidance or artifacts.
Key facts: - Modern AI coding emphasizes correctness via tests, agentic loops, and real-repo evaluation. - Reasoning-first and MoE approaches improve capability-per-compute when paired with tools. Design guidance with risks, metrics, acceptance criteria
{ "target_language": "Rust", "developer_needs": [ "cost_latency_tradeoffs", "evaluation_metrics", "ci_integration" ] }