Qwen3-4B-C-Coder-SFT-v1/README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-4B-Instruct-2507
pipeline_tag: text-generation
library_name: transformers
language:
  - en
tags:
  - code
  - c
  - qlora
  - unsloth
  - text-generation-inference
---

# Qwen3-4B-C-Coder-SFT-v1

A C-language coding specialist built on **Qwen3-4B-Instruct-2507**, fine-tuned with QLoRA
(rank 32, all linear layers) on ~23k curated C instruction pairs. This is **SFT v1** — the
first stage of a larger pipeline (synthetic execution-filtered data and GRPO with a
compiler/sanitizer reward are planned follow-ups).

The headline improvement is **instruction compliance for C**: the model follows exact
function signatures and emits compilable code far more reliably than its base.

## Results

All benchmarks are execution-based: generated C is compiled with gcc and run against hidden
tests in a sandboxed container. `safe-pass@1` additionally requires zero ASan/UBSan reports.

![CEval-priv: base vs SFT v1](images/ceval_priv.png)

| Benchmark | Metric | Base Qwen3-4B | **This model** | Δ |
|---|---|---|---|---|
| CEval-priv (161 tasks) | pass@1 | 36.6% | **59.0%** | **+22.4** |
| CEval-priv | compile rate | 41% | **80%** | +39 |
| McEval-C (50 tasks) | pass@1 | 52.0% | 48.0% | −4 (within ±14 pt CI) |
| McEval-C | compile rate | 72% | **90%** | +18 |

- **CEval-priv** is a private, contamination-proof eval set: 161 tasks machine-translated from
  HumanEval+/MBPP+ to C with deterministic type mapping, kept only if the reference solution
  compiles, passes its own tests, and runs sanitizer-clean. It was never trained on and all
  training data was decontaminated against it (10-gram overlap).
- **McEval-C** is the public McEval C-generation split. The pass@1 delta is within noise at
  n=50, single-sample; the compile-rate gain is the real signal.
- Sanitizer-clean pass rates equalled pass@1 for both models on both benchmarks.

**Interpretation:** one epoch of public C data (StackOverflow Q&A, curated instruction sets)
teaches *behavior* — signature compliance, compilable output — not new algorithmic ability.
Exactly what you'd expect, and what the later pipeline stages are for.

## Training

| | |
|---|---|
| Base | Qwen/Qwen3-4B-Instruct-2507 (non-thinking) |
| Method | QLoRA via Unsloth: r=32, α=64, dropout 0.05, all linear layers (66M trainable, 1.62%) |
| Data | 22,913 train / 467 valid; max seq 2048; loss on assistant tokens only |
| Schedule | 2,400 steps (≈0.84 epoch), effective batch 8, cosine LR 1e-4, warmup 100 |
| Hardware | Single Kaggle T4 (fp16), 8h33m, total compute cost $0 |
| Dynamics | train loss 1.78 → 0.97; eval loss 1.094 → 1.067, improving monotonically (no overfit) |

### Data

~34.7k cleaned pairs, sampled to 22.9k train after mixing:

| Slice | Count | Source / license |
|---|---|---|
| StackOverflow C Q&A | ~15k (capped) | [Mxode/StackOverflow-QA-C-Language-40k](https://huggingface.co/datasets/Mxode/StackOverflow-QA-C-Language-40k) — **CC BY-SA**; content © original Stack Overflow contributors, attribution per SO terms |
| Glaive code assistant v3 (C, syntax-verified) | ~3.9k | [glaiveai/glaive-code-assistant-v3](https://huggingface.co/datasets/glaiveai/glaive-code-assistant-v3) (Apache 2.0) — kept only samples whose code parses as C |
| McEval-Instruct (C) | ~1k | [Multilingual-Multimodal-NLP/McEval-Instruct](https://huggingface.co/datasets/Multilingual-Multimodal-NLP/McEval-Instruct) |
| General replay (anti-forgetting) | 15% | [allenai/tulu-3-sft-mixture](https://huggingface.co/datasets/allenai/tulu-3-sft-mixture) (ODC-BY) |

All slices were exact-deduplicated and decontaminated (word-level 10-gram overlap) against
McEval, MdEval, HumanEval(+), MBPP(+), and the private eval set.

## Usage

Non-thinking model — use Qwen's recommended sampling: `temperature=0.7, top_p=0.8, top_k=20`.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "harshpreet931/Qwen3-4B-C-Coder-SFT-v1"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

messages = [{"role": "user", "content":
    "Write a C function `int popcount32(unsigned int x)` that counts set bits."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=512, temperature=0.7, top_p=0.8, top_k=20)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))
```

Apple Silicon (4-bit MLX): see
[harshpreet931/Qwen3-4B-C-Coder-SFT-v1-mlx-4bit](https://huggingface.co/harshpreet931/Qwen3-4B-C-Coder-SFT-v1-mlx-4bit).
LoRA adapters only: [harshpreet931/Qwen3-4B-C-Coder-SFT-v1-LoRA](https://huggingface.co/harshpreet931/Qwen3-4B-C-Coder-SFT-v1-LoRA).

## Limitations

- **Algorithmic ability is unchanged from the base model.** This stage improved format/signature
  compliance and compile rates, not problem-solving. Don't expect gains on hard competitive tasks.
- English-only instruction data; C17/glibc-flavored; not tuned for embedded/kernel dialects.
- Inherits base-model limitations and possible biases; generated code should be reviewed and
  tested — compile-and-run verification (ideally with `-fsanitize=address,undefined`) is cheap, use it.

## Provenance

Built as part of an open, $0-compute project (MacBook M4 Pro + Kaggle free T4s): six-agent
research sweep → sandboxed compile/run/sanitizer harness → data pipeline → this SFT run.
Fun fact surfaced by the harness: one of McEval-C's own canonical solutions fails
LeakSanitizer.