Pharmaceutical domain preference dataset used for
Direct Preference Optimization (DPO) — Stage 3 of the pharma TinyLlama
fine-tuning pipeline.
Each JSONL record contains 3 fields:
{
"prompt": "### Instruction:\nExplain the mechanism of metformin.\n\n### Response:\n",
"chosen": "Metformin primarily works by ...",
"rejected": "Metformin is a drug that ..."
}
prompt — Alpaca-style instruction prompt (same format as… See the full description on the dataset page:
https://huggingface.co/datasets/ThakrePranjal/pharma-preference-dataset.
