---
license: mit
size_categories:
- n<1K
---

---
configs:   
  - config_name: raw   
    data_files: raw/pharma_paragraph_process.jsonl   
  - config_name: instruction   
    data_files: instruction/pharma_instruction_dataset.jsonl   
  - config_name: preference   
    data_files: preference/pharma_preference_dataset.jsonl   
---   

# Pharma Fine-Tuning Data

Three datasets used across the 3-stage fine-tuning pipeline in
[llm-finetuning-playbook](https://github.com/sivasaiyadav8143/llm-finetuning-playbook/tree/main):

- **raw**: cleaned paragraphs extracted from the pharma PDF (Stage 1 — continued pretraining)
- **instruction**: instruction/response pairs (Stage 2 — SFT)
- **preference**: prompt/chosen/rejected triples (Stage 3 — DPO)

Source corpus: Metformin pharmacology, statin/ezetimibe lipid-lowering therapy,
mRNA vaccines, AI in drug discovery, clinical trials & pharmacovigilance.

Educational project — not medical advice.