OxCoder-9B-GMCE (GMCoder Efficiency, experimental)
OrionLLM/OxCoder-9B fine-tuned for GMCoder Efficiency. Files: GGUF (Q8_0, Q6_K, Q5_K_M, Q4_K_M, BF16) and BF16 safetensors. The original checkpoint has no multi-token-prediction head, so the GGUFs are built with --no-mtp.
What it does (measured)
Our own unseen "Form B" exams, 60k token cap, single run per setting, graded blind by one grader. Scores are approximate.
| Setting | Historical | Everyday (/150) | Notes |
|---|---|---|---|
| OxCoder-9B, thinking on | looped to 60k, no answer | ~118 at 6.9k tokens | Ultimate ran to the cap |
| OxCoder-9B, thinking off | ~66 at 3.8k | 6.2k tokens (not graded) | |
| GMCE, thinking on | ~62 at 5.5k | ~105 at 5.9k | finishes Historical where the base looped; Ultimate ran to the cap |
| GMCE, thinking off | ~52 at 2.0k | 2.3k tokens (not graded) | all 3 exams in 5.7k tokens; Ultimate 11.5/25; invented citations |
Caveats
- Mixed result: GMCE stops OxCoder's runaway thinking on one exam, but costs about 13 points on the other. The base model with thinking off was a better cheap option in our tests.
- Published as an experiment. Benchmarks are our own and small.