Ternary quantization
Quantize Qwen3.5-4B to ternary weights with minimal accuracy loss
PrismMLInstruction
Quantize Qwen3.5-4B in /artifacts/qwen3.5-4b/ for text generation only. More than 95% of text model parameters outside the input embedding and output head must be ternary, with roughly one ternary code per original weight (stored parameters may exceed the original count by at most 10%); the embedding, head and activations may remain floating point. You may keep the vision tower and the MTP head in floating point as well, since they will not be used or evaluated.
Write serve.sh to serve an OpenAI-compatible chat endpoint offline on one H100, ready within 300 seconds.
Try to preserve as much performance as possible.
Do not train, fine-tune, calibrate or select on the validation or test split of any public dataset or benchmark; training splits and other public corpora are fine.
Submit in /workspace/output/. Interface and runtime details: /task/CONTRACT.md.
Supporting files: /task/public/. Evaluate with python /task/validate.py --server-url URL.
Test the finished submission the way it will be run with frontier selfcheck.
Hand it in with frontier submit when it is final.
Use spin --help for compute and model APIs.
Trials
42 graded · click a score for the trajectoryScoring
The raw metric (accuracy on GPQA Diamond, MATH-500, LiveCodeBench v6 and MMLU-Pro) is rescaled so that chance on each benchmark scores 0 and the original bf16 Qwen3.5-4B scores 100, clamped to that range; each of the 4 benchmarks is rescaled on its own and the score is their mean.
| Benchmark | Scores 0 at | Scores 100 at |
|---|---|---|
| gpqa_diamond | 0.25 | 0.5707 |
| math_500 | 0 | 0.866 |
| livecodebench_v6 | 0 | 0.4246 |
| mmlu_pro | 0.1111 | 0.7164 |