Model Summary: Mify-Coder-2.5B

Overview

Mify-Coder-2.5B-v1 is a breakthrough 2.5B-parameter code model fully designed, engineered, and trained at Infosys on 4.2T tokens on Mify-2.5B base model. Despite its compact size, Mify-Coder-2.5B-v1 sets a new benchmark for small language models, achieving performance parity with frontier open-source models in code generation and tool calling, along with exemplary performance on safety metrics in helpfulness and harmlessness, and superior throughput that surpasses larger frontier models.

Developed by: Infosys Ltd.

---

Architecture & Training

• Base Model: Mify-2.5B
• Training Phases:
  • Continual Pretraining (CPT): Next-token prediction with Fill-in-the-Middle (FIM) for structural infilling.
  • Supervised Fine-Tuning (SFT): Instruction alignment for coding tasks, function calling, and safety.
• Optimization:
  • BF16 mixed precision, Grouped Query Attention (GQA), and Distributed Fused Adam optimizer.
  • Specialized tokenization with syntax markers and reasoning tokens for advanced behaviors.

---

Performance Highlights

Category	Benchmark	# Shots	Metric	Scores
Code Gen	MBPP	0	pass@1	91.21%
Code Gen	MBPP+	0	pass@1	89.15%
Code Gen	HumanEval	0	pass@1	53.66%
Code Gen	HumanEval+	0	pass@1	48.78%
Code Gen	NumpyEval	0	pass@1	56.44%
Code Gen	PandasEval	0	pass@1	53.47%
Tool Use	BFCL v2	0	overall acc	55.26%
Safety	AIR-Bench	0	pass@1	67.32%
SecCode Gen	CybersecEval4-Autocomplete	0	pass@1	78.91%

---

Responsible AI & Safety

• Integrated safety objectives during SFT.
• Balanced harmful/general sample ratio (1:4) for secure code generation and ethical language use.
• Validated against Stanford AIR-Bench and CybersecEval4-Autocomplete benchmarks.

---

Deployment & Future Work

• Quantization: The model was optimized for low latency outperforming most sub-8B SLM models. Furthermore, the quantized variants of Mify-Coder can be seamlessly deployed and inferenced on standard desktop environments, eliminating the need for specialized hardware such as GPUs.
• Future work includes enhancing Mify-Coder with agentic coding competencies and scaling its context length. The model weights will be open-sourced early next year to accelerate research and real-world deployment.