---
base_model:
- Qwen/Qwen2.5-Coder-3B-Instruct
library_name: peft
pipeline_tag: text-generation
tags:
- lora
- sft
- transformers
- unsloth
- qlora
- macaulay2
- peft
- code-generation
license: mit
datasets:
- frupniew/macaulay2-qa-instruct
language:
- en


# Macaulay2 RAG-Coder 3B (LoRA Adapter)

This repository contains the **QLoRA adapter weights** trained to turn `Qwen2.5-Coder-3B-Instruct` into a domain expert for Macaulay2 (commutative algebra). 

## 🛠️ Training Details & Post-Mortem
- **Framework:** [Unsloth](https://github.com/unslothai/unsloth) (QLoRA 4-bit)
- **Hyperparameters:** `r=16`, `alpha=32`, `lr=2e-4`, `epochs=5`, `packing=True` (max_seq_len=4096).
- **Target Modules:** All linear layers.
- **Engineering Post-Mortem (Task Mismatch):** Initial training resulted in a model that perfectly understood M2 syntax but wrapped outputs in JSON (due to the synthetic data generator's system prompt). A **continual learning retrain** (Path A) was executed on pure ChatML formatted data to force raw Markdown code blocks, successfully resolving the issue.
- **PEFT Quirk Handled:** Addressed the `tie_word_embeddings` issue (PEFT #2777) specific to Qwen2.5 architecture during the merge process to prevent generation degradation.

## 🐍 Usage with Transformers
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "Qwen/Qwen2.5-Coder-3B-Instruct"
adapter_model_id = "frupniew/macaulay2-rag-coder-3b-adapter"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(base_model_id, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base_model, adapter_model_id)

prompt = "How do I compute the primary decomposition of an ideal?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))