ThakrePranjal/pharma-tinyllama-unsloth-final

Pharma TinyLlama — Final Model (Unsloth 3-Stage Complete)

This is the final fully merged standalone model — complete output of the Unsloth 3-stage pharma fine-tuning pipeline applied to unsloth/tinyllama-bnb-4bit.

3-Stage Pipeline Summary

Stage	Type	Data	What model learned
1	Domain Pretraining (SFT)	Raw pharma PDF paragraphs	Pharma vocabulary, facts, terminology
2	Instruction Fine-Tuning (SFT)	48 Alpaca-format QA pairs	How to follow pharma instructions
3	Preference Tuning (DPO)	48 chosen/rejected pairs	Which pharma answers are better

Built with Unsloth (2x faster, lower VRAM) + HuggingFace TRL. Total GPU time: ~353 seconds on Tesla T4.

Quick Start

from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="ThakrePranjal/pharma-tinyllama-unsloth-final",
    max_seq_length=512,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

def ask(question, max_new_tokens=150):
    prompt = f"### Instruction:\n{question}\n\n### Response:\n"
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    with torch.inference_mode():
        out = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            do_sample=True,
            temperature=0.7,
            top_p=0.9,
            repetition_penalty=1.1,
            pad_token_id=tokenizer.eos_token_id,
        )
    input_len = inputs["input_ids"].shape[-1]
    return tokenizer.decode(out[0][input_len:], skip_special_tokens=True).strip()

# Test
questions = [
    "Explain the primary mechanism of action of metformin.",
    "Why should AI predictions in drug discovery be experimentally validated?",
    "Define pharmacovigilance.",
    "Why should atorvastatin and ezetimibe be used together?",
]
for q in questions:
    print(f"Q: {q}")
    print(f"A: {ask(q)}\n")

Complete Pipeline Repos

Stage	Dataset	LoRA Adapter	Merged Model
1 — Domain Pretraining	—	ThakrePranjal/pharma-tinyllama-unsloth-stage1-lora	ThakrePranjal/pharma-tinyllama-unsloth-stage1-merged
2 — Instruction SFT	ThakrePranjal/pharma-instruction-dataset-unsloth	ThakrePranjal/pharma-tinyllama-unsloth-stage2-lora	ThakrePranjal/pharma-tinyllama-unsloth-stage2-merged
3 — DPO Preference	ThakrePranjal/pharma-preference-dataset-unsloth	ThakrePranjal/pharma-tinyllama-unsloth-stage3-dpo-lora	THIS MODEL

Limitations

Trained on a small pharma corpus (1 PDF, 48 instruction/preference pairs). Not validated for clinical or production use. Intended for educational/research purposes only.