drugenv-qwen25-05b-lora

DrugEnv LoRA Adapter (Qwen2.5-0.5B)

This repo contains a LoRA adapter only fine-tuned on a small synthetic instruction dataset to improve DrugEnv action JSON formatting (especially parameter schema consistency) for the Drug Target Validation environment.

• Base model: Qwen/Qwen2.5-0.5B-Instruct
• Adapter: LoRA (adapter_model.safetensors)
• Training: ~300 steps on free Colab T4, 4-bit quantized base model

What Changed (Quick Example)

Before (base model tends to emit non-env parameter keys):

{"action_type":"query_expression","parameters":{"term":"EGFR","dossier_id":"NSCLC"},"reasoning":"..."}

After (base + adapter tends to emit env-aligned payloads):

{"action_type":"query_expression","parameters":{"database":"GTEx"},"reasoning":"..."}

How To Use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

base_id = "Qwen/Qwen2.5-0.5B-Instruct"
adapter_id = "anugrahteesdollar/drugenv-qwen25-05b-lora"

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.float16,
)

tok = AutoTokenizer.from_pretrained(base_id, use_fast=True)
if tok.pad_token is None:
    tok.pad_token = tok.eos_token

base = AutoModelForCausalLM.from_pretrained(
    base_id, device_map="auto", quantization_config=bnb, torch_dtype=torch.float16
)
model = PeftModel.from_pretrained(base, adapter_id)

## Training Data

Small synthetic instruction dataset embedded directly in a Colab notebook (focused on emitting valid JSON actions for DrugEnv).

## Limitations

This is a hackathon-speed adapter meant to demonstrate formatting/behavior improvements quickly; it is not a full GRPO-trained
policy.


Notes:
- Remove the invalid tag `base_model:adapter:...` (it can confuse HF parsing).
- The “Framework versions” line doesn’t need to be in YAML; you can mention it in text if you want. If you do, add a short line
near the end: “Trained with PEFT 0.19.1.”