aurelius-14b

Aurelius-14B

A clean, verification-native code fine-tune of Qwen/Qwen3-14B. Supervised on execution-verified solutions distilled from Qwen3-Coder-30B-A3B-Instruct (Apache-2.0) — no closed-model outputs. Designed to be paired with a verifier-native best-of-N + repair serving layer.

Results

Held-out HumanEval is the clean judge (training data is MBPP-distilled; the two sets are disjoint). Single-pass ≈ base — SFT here is safe, non-regressing polish, not a capability lift. The large gain is the serving layer, which cashes the model's selection gap (oracle@32 = 97.0%).

Honest methodology — what moved the needle, and what didn't

The negatives are reported deliberately: on this base, single-pass capability is bounded by the base model; the cheap, reliable gain lives in verifier-native inference, not in more post-training.

Provenance & license

• Base: Qwen3-14B (Apache-2.0). Teacher: Qwen3-Coder-30B-A3B-Instruct (Apache-2.0).
• Training data: best-of-N MBPP solutions, kept only if they pass the unit tests.
• 100% open-weight provenance — no Claude / GPT / closed-model outputs.
• License: Apache-2.0 (inherits base + teacher).

Usage

Single-pass:

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Zephyrs33/aurelius-14b")
model = AutoModelForCausalLM.from_pretrained("Zephyrs33/aurelius-14b", device_map="auto")
msg = [{"role": "user", "content": "Write a Python function is_prime(n). Return ONLY the function in one ```python block."}]
ids = tok.apply_chat_template(msg, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=512)[0][ids.shape[1]:], skip_special_tokens=True))

For the best-of-N + repair serving layer (the 96.3% / 83.0% numbers), use solve(problem, tests) — sample N, execute against a verifier, return a passing solution; repair on failure.

Limitations

Code-focused (MBPP / HumanEval distribution). Single-pass capability ≈ the base; the serving lift requires a verifier (your tests / sample tests / generated tests — never hidden eval tests). Not safety-tuned beyond the base.