clinicalabstainsql-artifacts

Large artifacts for the BN5212 project ClinicalAbstain-Text2SQL-LLM (code, report and instructions: https://github.com/nitinvetcha/ClinicalAbstain-Text2SQL-LLM — see its README and repro/REPORT.md).

All paths mirror the GitHub repo's layout, so download into the repo root.

Path	What it is
repro/runs/lora_seed42/	Llama-3-8B-Instruct LoRA adapter (answerable + unanswerable data = abstention-aware). Final adapter at the top level, plus checkpoint-150/300/309.
repro/runs/lora_sqlonly_seed42/	Same, trained on answerable (SQL-only) data. Checkpoints 150/279.
repro/multi_model/Qwen3-8B/{run,run_sqlonly}/	Same two variants on Qwen3-8B.
repro/multi_model/Qwen3-4B-Instruct-2507/{run,run_sqlonly}/	Same on Qwen3-4B-Instruct-2507.
repro/multi_model/Mistral-7B-Instruct-v0.3/{run,run_sqlonly}/	Same on Mistral-7B-Instruct-v0.3.
repro/pgdata_synthetic.tar.gz	PostgreSQL data directory of the synthetic OMOP benchmark DB (tar xzf inside repro/).

Each adapter dir has adapter_model.safetensors, adapter_config.json, tokenizer files, train.log, eval_results.json.

Training data: only the synthetic OMOP text-to-SQL benchmark (Legrand et al., https://github.com/red1dwarf/TEXT-TO-OMOP). No MIMIC-IV / MIMIC-IV-Note data or MIMIC-derived files are included (PhysioNet credentialed data; get your own access). These are adapters only; load them on top of the matching base model (separately licensed: Meta Llama-3, Qwen3, Mistral), e.g.

huggingface-cli download Nitin2004/clinicalabstainsql-artifacts --include "repro/runs/lora_seed42/*" --local-dir .