qwen-2.5-construction-codecite-v1

Qwen 2.5 1.5B — Construction Code-Citation v1

LoRA adapter on top of Qwen 2.5 1.5B-Instruct (4-bit MLX). Predicts OIICS hazard codes (event, source, nature, body) and OSHA 29 CFR 1926 citations from construction-site incident narratives.

Built for the Adaption Labs AutoScientist Challenge ("All Other Domains" category).

Inputs / Outputs

Input: free-text construction-site narrative. Output: strict JSON with hazards[] (4 OIICS codes + severity) and citations[] (verified OSHA 1926 standards).

Usage

from mlx_lm import load, generate

model, tokenizer = load(
    "mlx-community/Qwen2.5-1.5B-Instruct-4bit",
    adapter_path="oversite/qwen-2.5-construction-codecite-v1",
)
prompt = "Worker fell from second-story scaffold..."
out = generate(model, tokenizer, prompt=prompt, max_tokens=384)

See gradio_app/app.py in the source repo for the full prompt template and RAG-augmented inference pipeline.

Training

• Base model: Qwen 2.5 1.5B-Instruct, 4-bit MLX quantization
• Method: LoRA, 16 layers, 5.3M trainable parameters
• Data: 17,127 stratified-by-event-division SFT examples from OSHA SIR (2015-2025)
• Optimizer: Adam, lr 1e-4, batch 2, 400 iterations
• Loss masking: prompt masked, train on completion tokens only
• Seed: 20260606

Metrics

n=200, split=dev, git_sha=d926d81

Test-set numbers are held back until submission per the locked split (SHA-256 c9490ed3...).

Limitations

• Source-code distribution has a heavy long tail (1,478 unique codes). Model uses an OTHER bucket for codes outside the top-75 shortlist.
• Citation grounding is BM25-only at v1 (vector index follow-up).
• SIR over-represents severe injuries; the model is biased toward high-severity event types.

License

MIT. Base model Qwen 2.5 1.5B-Instruct is governed by its upstream license.

Citation

@misc{construction-code-llm-2026,
  title  = {Qwen 2.5 1.5B - Construction Code-Citation v1},
  author = {Oversite Innovations},
  year   = {2026}
}