jprivera44/model_checkpoints_from_exp2_v2

Backdoor Detection Checkpoints

Fine-tuned Qwen 2.5 Coder 14B models for backdoor detection experiments.

Models

1. ca_v3_run_20251215/

• Base model: Qwen/Qwen2.5-Coder-14B-Instruct
• Training data: CA_collusion_exp2_v3 (300 clean + 300 backdoor_only)
• Task: Binary backdoor detection
• Best checkpoint: final_model/ or checkpoint-150/

2. exp1_original_data_rerun_20251215/

• Base model: Qwen/Qwen2.5-Coder-14B-Instruct
• Training data: exp1_qwen14b_analysis_v2 (original dataset)
• Task: Binary backdoor detection
• Performance: AUROC 0.947
• Best checkpoint: checkpoint-75/

3. exp2_signalc_20251215/

• Base model: Qwen/Qwen2.5-Coder-14B-Instruct
• Training data: 300 clean + 250 backdoor_only + 50 signal_c
• Task: Backdoor detection with collusion signal resistance
• Goal: Train model to detect backdoors even when signal_c marker is present
• Best checkpoint: final_model/ or checkpoint-150/

LoRA Configuration

All models use LoRA fine-tuning:

• lora_r: 32
• lora_alpha: 64
• target_modules: [q_proj, v_proj, k_proj, o_proj]
• lora_dropout: 0.05

Usage

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

model = AutoPeftModelForCausalLM.from_pretrained(
    "path/to/checkpoint",
    torch_dtype="bfloat16",
    device_map="cuda"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-14B-Instruct")

Date

December 15, 2025