incidentbench-silver_human

LoRA adapter (r=16) on Qwen2.5-7B-Instruct for the IncidentBench joint classifier: one answer from seven MIT risk domains or No incident (potential only). Detection and category are both read from that one answer.

Question this arm answers: Silver_train with human_train labels replacing the silver label on its rows (alleged = not an incident).

Training labels are model-written ("silver") under the project's annotation guideline, not human gold. Research artefact: scores and known failure modes (it under-calls incidents) are in the IncidentBench report; do not use it for decisions about real systems.

Prompt format, training rows, and hashes are in data_manifest.json; the training recipe is in train_manifest.json. The adapter is applied unmerged on top of the base model (as in every reported run). Serve with vLLM: vllm serve Qwen/Qwen2.5-7B-Instruct --enable-lora --max-lora-rank 16 --lora-modules silver_human=chiffonng/incidentbench-silver_human and send the arm's prompt (train/build_data.py transform orig), greedy decoding. Training stack: transformers 5.9.0, peft 0.19.1.