README.md

---
language:
- en
- code
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
tags:
- lora
- code
- qwen2.5-coder
- fingpt
- code-correction
pipeline_tag: text-generation
---

# fingpt-coder-1b5

LoRA adapter for **[Qwen/Qwen2.5-Coder-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct)** fine-tuned on
[m-a-p/Code-Feedback](https://huggingface.co/datasets/m-a-p/Code-Feedback)
(66K error→fix pairs, 3 epochs).

> **Adapter only** — the base model is loaded from the HF Hub automatically.
> Total download: ~84 MB adapter + ~3 GB base model.

---

## LoRA config

| Property | Value |
|----------|-------|
| Base model | `Qwen/Qwen2.5-Coder-1.5B-Instruct` |
| Rank (r) | 16 |
| Alpha | 32 (scale = 2.0) |
| Target modules | `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj` |
| Training step | 48500 |
| Adapter size | ~84 MB |

---

## Quick start

```bash
git clone https://huggingface.co/revana/fingpt-coder-1b5
```

```python
import torch, sys
sys.path.insert(0, "fingpt")          # fingpt repo root
from infer import load_model, generate

model, tokenizer = load_model("adapter_final.pt")

reply = generate(model, tokenizer, "Fix this bug:\n\ndef fact(n):\n    return n * fact(n)")
print(reply)
```

Or use the [live demo](https://huggingface.co/spaces/revana/fingpt).

---

## Training

| Property | Value |
|----------|-------|
| Dataset | [m-a-p/Code-Feedback](https://huggingface.co/datasets/m-a-p/Code-Feedback) |
| Samples | ~66K error→fix pairs |
| Epochs | 3 |
| Batch size | 4 × 4 grad accum = 16 effective |
| LR | 3e-4, cosine decay, 3% warmup |
| Precision | bfloat16 |
| Hardware | A100 80GB |

---

## License

Apache 2.0