North-star: draft-only router adapters for North-Mini-Code-1.0
Research artifacts, not a deployable model. This repo holds two small draft-only router adapters (low-rank, rank 4, 100 training steps) produced by the weight-edit track of Bobliuuu/north-star. No North weights are included or redistributed; the adapters are only meaningful on top of CohereLabs/North-Mini-Code-1.0 (revision d11e61a842617a22dc328552fa5bb86231ee4f37), and their use is subject to that model's license. The adapter files are derivatives of the base model's routers, so the base model's license terms apply.
What they do
An adapter adds a low-rank correction to the router logits only while drafting (top-2 experts kept). The verifier path is untouched: verifier logits are bit-identical with the adapter installed and the base parameters are unchanged, so speculative output stays exactly the verifier's greedy output.
| v001 | v002 (CURRENT) | |
|---|---|---|
| adapter params | 383,020 | 348,200 |
| draft top-1 agreement, base -> edited | 0.654 -> 0.702 | 0.454 -> 0.571 |
| committed tokens per block (gamma=4), base -> edited | 2.71 -> 2.93 | 1.84 -> 2.31 |
| gates (lossless, gain, overfit, cost) | all pass | all pass |
| adapter FLOPs fraction | 0.0009 | 0.0009 |
Measured on one H100 80GB, 8 gate sequences, sampled mode; numbers are from v00*/report.json. These are offline draft-agreement numbers, not end-to-end tok/s speedups; no end-to-end speedup from these adapters is claimed.
Files
v001/, v002/: adapters.pt (a dict of per-layer A.* / B.* router tensors), recipe.json, report.json. Load with the weight-edit code in the GitHub repo (src/weightedit/).
Links
- Code, results, write-ups: https://github.com/Bobliuuu/north-star (see
docs/WEIGHT_EDIT.md,RESULTS.md) - Explainer video: https://github.com/Bobliuuu/north-star/blob/main/video/northstar_explainer.mp4
- Release: https://github.com/Bobliuuu/north-star/releases/tag/v0.2.0
Model tree for Bobliu/north-star
Base model
CohereLabs/North-Mini-Code-1.0