---
license: apache-2.0
base_model: Qwen/Qwen3.5-0.8B-Base
tags:
  - qwen
  - python
  - code-generation
  - qlora
  - sft
  - unsloth
  - gguf
datasets:
  - code-search-net/code_search_net
language:
  - en
library_name: transformers
pipeline_tag: text-generation
---

# Qwen3.5-0.8B-Python-SFT

**Python code generation model** — Qwen3.5-0.8B-Base fine-tuned with **QLoRA
(Supervised Fine-Tuning)** on **CodeSearchNet (Python)**: docstring → function
code pairs.

## Model Details

| Property | Value |
|----------|-------|
| Base model | [Qwen/Qwen3.5-0.8B-Base](https://huggingface.co/Qwen/Qwen3.5-0.8B-Base) |
| Method | QLoRA (4-bit base + LoRA r=16, alpha=32) |
| Trainable params | 6.4M / 759M (0.84%) |
| Dataset | CodeSearchNet Python — 408K samples (13,590 repos) |
| Task | Docstring → Python function code |
| Sequence length | 2048 |
| Precision | BF16 |
| Hardware | NVIDIA RTX 4060 8GB |

## Training Results

| Metric | Value |
|--------|-------|
| Train loss | 0.330 |
| Eval loss | 1.214 |
| Steps | 25,524 (1 epoch) |
| Runtime | ~25.7h |

## Evaluation

pass@1 (temperature 0.2), official test harness, both models in bf16.

| Benchmark | Base | **Fine-tuned** | Improvement |
|-----------|-----:|---------------:|------------:|
| HumanEval | 1.2% | **17.7%** | 14.5x |
| MBPP | 0.0% | 0.2% | 0 → 1 |

Full report with example solutions: `reports/evaluation_report.md` in the training repo.

## Training Details

- **Method:** QLoRA — 4-bit quantized base + LoRA (r=16, alpha=32, dropout=0)
- **Target modules:** q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
- **Optimizer:** adamw_8bit (bitsandbytes), cosine schedule, 3% warmup
- **Batch:** 2 per device × 8 grad accumulation (effective 16)
- **Max sequence length:** 2048
- **Hardware:** NVIDIA RTX 4060 8GB, ~25.7h
- **Data:** CodeSearchNet Python filtered to ≤2048 tokens (408,377 train samples)

## Usage

### Transformers (LoRA adapter)

```python
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "Qwen/Qwen3.5-0.8B-Base",
    max_seq_length=2048,
    load_in_4bit=True,
)
model, tokenizer = FastLanguageModel.from_pretrained(
    "jaweed123/Qwen3.5-0.8B-Python-SFT",
    max_seq_length=2048,
    load_in_4bit=True,
)
```

### GGUF (ollama / llama.cpp / vLLM)

```bash
# llama.cpp
llama-cli -m qwen3.5-0.8b-python-sft-q4_k_m.gguf -p "Write a Python function that..."

# Ollama
ollama create qwen3.5-python -f Modelfile
```

```dockerfile
# Modelfile
FROM qwen3.5-0.8b-python-sft-q4_k_m.gguf
TEMPLATE "{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
"
```

## Files

| File | Description |
|------|-------------|
| `adapter_model.safetensors` | LoRA adapter (small, ~13MB) |
| `model.safetensors` | Merged 16-bit model |
| `qwen3.5-0.8b-python-sft-q4_k_m.gguf` | GGUF Q4_K_M (~0.5GB) |
| `qwen3.5-0.8b-python-sft-q8_0.gguf` | GGUF Q8_0 (~0.9GB) |
| `qwen3.5-0.8b-python-sft-f16.gguf` | GGUF F16 |

## Limitations

- Fine-tuned for Python function generation from docstrings
- Trained on 2019-era open-source code
- 0.8B scale — limited reasoning; best for straightforward code tasks