bonsai: depth24/experts0
Pruned and distilled from Qwen3-30B-A3B-Base for TypeScript code completion. Part of the bonsai project, which applies the Minitron structured-pruning and knowledge-distillation recipe (Muralidharan et al. 2024, Sreenivas et al. 2024) to a Mixture-of-Experts model, a pruning axis neither paper covers.
What was pruned
- Depth: 24 of 48 layers removed (24 kept)
- Experts: 0 of 128 routed experts removed per layer (128 kept)
- Result: ~15.58B parameters (51.1% of baseline's ~30.5B)
Training
500-step fixed-loss short schedule (75-step warmup, peak LR 1e-4), distilled against the uncorrected teacher. Best = final checkpoint (step 499), no earlier minimum.
Results
HumanEval-TS (TypeScript MultiPL-E), full 159-problem suite, temperature 0.2, top-p 0.95:
| pass@1 | |
|---|---|
| Qwen3-30B-A3B-Base (unpruned) | 0.717 (114/159) |
| This model | 0.107 (17/159) |
14.9% of baseline accuracy at 51.1% of baseline's parameters.
Notes
Pure depth-only cut, half the model's layers removed. Costs far more accuracy per parameter removed than an equivalently-sized experts-only cut (depth0/experts64, same ballpark size, 5.4x the accuracy). Not recommended. See depth0/experts64.
Project
Other checkpoints from this sweep: bonsai collection. Code and full writeup: github.com/SartajBhuvaji/bonsai (private repo).