Winnow
Collection
Variable-width MoE expert pruning (Winnow, formerly GLEAN): pruned OLMoE, Qwen3.6-35B-A3B and poolside Laguna-S-2.1 checkpoints, unhealed and healed. • 15 items • Updated
winnow-olmoe-math-keep25
Channel-level pruned + healed OLMoE-1B-7B-0125-Instruct.
Requires trust_remote_code=True (ragged variable-width experts).
| Base | allenai/OLMoE-1B-7B-0125-Instruct |
| Params | 2.09B |
| Experts fully deleted | 442/1024 |
| Keep fraction | 0.25 |
| Criterion | channel-level REAP, per-layer budgets, block 128, min width 128 |
| Calibration | Dolmino-math (scores_0125inst_dolmino-math) |
| Final forward top-128 KL | 0.088 |
Healing
Off-policy forward-KL distillation against cached top-128 teacher targets
(dolci_math_curated_opd_top128), 150 steps, 120k loss tokens/step (18M total), AdamW8bit,
lr 3e-5, wd 0.1, 10 warmup steps, grad clip 1.0, chat frames, max seq len 2048, seed 1223.
All six models in this release share this recipe and are matched on optimizer
steps and loss tokens. Note they are not matched on data exposure: the math
cache is 6.48M unique tokens (2.8 epochs at 150 steps) while the general cache
is 22.3M (0.81 epochs).
- Downloads last month
- 7
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for hbfreed/winnow-olmoe-math-keep25
Base model
allenai/OLMoE-1B-7B-0125 Finetuned
allenai/OLMoE-1B-7B-0125-SFT Finetuned
allenai/OLMoE-1B-7B-0125-DPO Finetuned
allenai/OLMoE-1B-7B-0125-Instruct