ThoxEdge-RustCoder-0.5B

ThoxEdge-RustCoder-0.5B

A Rust-leaning fine-tune of Qwen/Qwen2.5-Coder-0.5B-Instruct, trained by THOX.ai on the Apache-2.0 Fortytwo-Network/Strandset-Rust-v1 corpus.

Both the base and this model are Apache-2.0. Nothing here is derived from, initialised from, or distilled out of any Gemma-licensed Rust model — see the lane doc for why that mattered.

What it is, and what it is not

It is a 0.5B model. It is not a strong Rust coder, and no number below should be read as claiming otherwise. It exists to serve cheap Rust completion in ThoxRoute's lowest tier and to establish whether the recipe moves the needle at all before the same recipe is spent on a 1.5B.

It does not fit THOX's 512 MB devices. The GGUF ladder for this architecture spans only 260 MB — Q2_K still costs 415 MB — because vocab_size 151,936 against hidden_size 896 puts 27.6% of params in an embedding whose row length is not divisible by the 256-wide K-quant super-block, so it never quantizes down. This is a hosted / desktop model.

Evaluation

MultiPL-E humaneval-rs, 156 problems, executed: each generation is assembled into a program, compiled with rustc, and its asserts run. Greedy decoding, so the delta is not sampling noise.

Both models were scored in the same job, so the comparison cannot drift between runs.

The fine-tune improves on its base in both prompting modes, and the best available configuration moves from 25.64% (base, chat) to 28.21% (tuned, chat) — a net +2.57 pp, or four more HumanEval-Rust problems solved.

Noise floor: about ±1 problem (±0.64 pp). An earlier run of the identical base on different hardware scored 20.51% in completion mode against 21.15% here, despite greedy decoding — GPU reductions are not bit-identical across flavors. The +4.49 pp (7 problems) and +2.57 pp (4 problems) deltas both clear that, but anything under about 1.5 pp from this harness should be treated as a tie.

Assembly is deliberately generous — a body-only answer and a whole-function answer are both attempted, and the problem passes if either works. These numbers are therefore not comparable to published MultiPL-E leaderboards. The protocol is identical for both models, so the delta between the rows is the measurement.

Training

The ~117k rows dropped emit English about Rust — docstrings, naming, summaries, review. Training them would spend the step budget on describing code rather than writing it.

Prompting it

Ask for a complete function body in plain terms — for example "Write the complete body of the Rust function below."

Do not prompt it with "Complete the Rust code. Output only the missing code." That phrasing was the code_completion instruction across 14,527 training rows whose targets are short fill-in-the-middle fragments, so the model will answer it with a fragment — sometimes a bare identifier. Measured: that exact wording scored 2.56% where the recommended wording scores 28.21%. The training instruction has since been reworded to name both sides of the gap, but this checkpoint predates that fix and retains the sensitivity.

Known limitations in this build

• Trained with a renderer that placed the rationale / bug_description output fields inside the ```rust fence, affecting ~53% of rows across the refactoring, optimization and bug-detection categories. Fixed in tools/train_thox_rustcoder.py after this run. It did not cause any measured regression — that was the prompt collision above — but a rerun should still carry the fix.
• It is a 0.5B. 28.21% pass@1 means it fails roughly seven of every ten HumanEval-Rust problems.

Attribution

Training data: Fortytwo-Network/Strandset-Rust-v1, Apache-2.0 — synthetic Rust generated from idioms mined across 2,300+ crates.io crates, consensus-filtered and 94.3% rustc-verified.