---
language:
- en
tags:
- pharma
- preference-tuning
- dpo
- rlhf
- chosen-rejected
- unsloth
pretty_name: Pharma DPO Preference Dataset (Unsloth Pipeline)
---

# Pharma DPO Preference Dataset — Unsloth Pipeline

Preference dataset in **DPO format** (prompt / chosen / rejected) used for
Stage 3 DPO training in the Unsloth 3-stage pharma fine-tuning pipeline.

## Format

```json
{
  "prompt":   "### Instruction:\nExplain the mechanism of metformin.\n\n### Response:",
  "chosen":   "Metformin primarily acts by activating AMPK...",
  "rejected": "Metformin mainly works by increasing insulin secretion..."
}
```

## Stats

- Total rows: 48
- Used as-is for DPO training (no validation split)

## DPO training details

| Param | Value |
|---|---|
| Beta (KL penalty) | 0.1 |
| Max steps | 30 |
| Learning rate | 5e-5 |
| Trainer | `trl.DPOTrainer` (Unsloth patched) |

## Pipeline context

```
Stage 1 (SFT - Domain Pretraining)  → raw PDF paragraphs
Stage 2 (SFT - Instruction Tuning)  → [ThakrePranjal/pharma-instruction-dataset-unsloth](https://huggingface.co/datasets/ThakrePranjal/pharma-instruction-dataset-unsloth)
Stage 3 (DPO - Preference Tuning)   → THIS DATASET
```

## Related models

- Stage 3 DPO LoRA : [ThakrePranjal/pharma-tinyllama-unsloth-stage3-dpo-lora](https://huggingface.co/ThakrePranjal/pharma-tinyllama-unsloth-stage3-dpo-lora)
- Final merged     : [ThakrePranjal/pharma-tinyllama-unsloth-final](https://huggingface.co/ThakrePranjal/pharma-tinyllama-unsloth-final)