Modelpublic

semianalysisai/Qwen3.6-35B-A3B-graph-color-GRPO-20260909

sourceHugging Faceupdated 18d agoView on Hugging Face

0likes28downloads

Model Card

Qwen3.6-35B-A3B · Graph Coloring

A specialist checkpoint for graph-coloring puzzles, trained from Qwen3.6-35B-A3B with 40 updates of verifier-reward GRPO (Group Relative Policy Optimization).

Training

Setting	Value
Base model	`Qwen/Qwen3.6-35B-A3B`
Training puzzles	10,000, generated with Reasoning Gym
GRPO updates	40
Batch per update	32 prompts × 8 responses = 256 responses
Learning rate	`1e-6`
Response limit	256 tokens; thinking disabled
Hardware	1 node with 8 × NVIDIA B200 GPUs
Training framework	Miles

Evaluation and scope

—Development evaluation: 512 puzzles per domain. Scores are not reported in this model card.
—Held-out test set: not evaluated.
—Export: trained language-model weights. Multimodal capabilities are not validated.
—Validate independently before downstream use. Equivalence to the results reported in the Miles PR has not been established.

<details> <summary><strong>Reproducibility and implementation details</strong></summary>

Each specialist was trained on its own puzzle domain. Reasoning Gym generated 10,000 training puzzles per domain.

Component	Source revision
Base model	`995ad96eacd98c81ed38be0c5b274b04031597b0`
Miles — `radixark/miles`, PR #3116	`8f8e4dff55d80bd9d36c3490c333a119f38480ed`
Reasoning Gym	`49b07130b3fcd12f2d064bba7c43869543a0e7e7`

SGLang used the Triton MoE backend because automatic backend selection failed during weight transfer.

</details>

Companion checkpoint: Countdown