btl-2-coder-7B

BTL-2 Coder 7B

BTL-2 Coder 7B is a LoRA adapter for unsloth/Qwen2.5-Coder-7B-Instruct, trained for structured code-review findings.

Intended Use

Use this model for local-first code review:

• SQL injection
• path traversal
• authorization bypass
• missing error handling
• boundary/off-by-one logic
• related security and correctness bugs

It is not yet a general autonomous coding agent and should not be marketed as a SWE-Bench repair model.

Training

• Base: unsloth/Qwen2.5-Coder-7B-Instruct
• Trainer: Unsloth LoRA SFT
• Data: 4,000 API teacher traces + 1,000 template traces
• Split: 4,500 train / 500 eval
• Epochs: 2
• Max length: 4096

Only redacted, opt-in traces should be used for future training.

Prompt

Use strict schema prompting:

Return only a JSON array. No markdown and no wrapper object.
Each finding must include: severity, file, line, title, evidence, recommendation, confidence.
severity must be exactly one of: critical, high, medium, low.
Never put a category in severity.
confidence must be a number from 0 to 1, never a string label.
Every finding must include concrete evidence and a non-empty recommendation.

Example output:

[
  {
    "severity": "critical",
    "file": "src/users.ts",
    "line": 42,
    "title": "SQL injection through string-built query",
    "evidence": "The user id is concatenated directly into the SQL string.",
    "recommendation": "Use a parameterized query.",
    "confidence": 0.96
  }
]

Evaluation

Limitations

• Strict schema prompting is required for best results.
• The model may miss subtle multi-file issues.
• The model can produce plausible but incorrect findings; keep human review in the loop.
• Do not use on private repositories unless you control the inference environment and data policy.

Release Artifacts

This Hugging Face repo should include:

• adapter_model.safetensors
• adapter_config.json
• tokenizer.json
• tokenizer_config.json
• chat_template.jinja
• training_args.bin
• this README.md