MedHal
Collection
MedHal: A Synthetic Dataset for Medical Hallucination Detection • 8 items • Updated
medhal-llama3-8b
LoRA fine-tune of meta-llama/Meta-Llama-3-8B-Instruct for medical
hallucination detection, trained for MedHal.
Given a context and a statement, the model judges whether the statement is
factually supported by the context.
Training
- Method: supervised fine-tuning with LoRA (r=16, 4-bit), 1 epoch,
max_seq_length=8192, per-device batch 2, gradient accumulation 2, seed 42. Loss computed only after the### Factualmarker. - Data:
medhal_lf_bal_qafil_train_val. This artifact is the merged full weights (base + adapter viasave_trained_model.py).
Use
Format prompts with the MedHal formatter (src/data/formatter.py,
model_wrapper=medhal) and parse YES/NO answers with
src/evaluation/medhal_parser.py:MedHalParser.
Limitations
Research artifact. Not for clinical use. Inherits the base model's license
and biases; hallucinations in the judge itself are possible — see the
invalid rate reported by MedHalParser.evaluate.
- Downloads last month
- 4
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for LAMA-WeST/medhal-llama3-8b
Base model
meta-llama/Meta-Llama-3-8B-Instruct