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.
Table with columns: Benchmark, Metric, Base Qwen3-4B, This model, Δ| 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
Table | |
|---|
| 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:
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.
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.
LoRA adapters only: 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.