qwen3.6-35b-a3b-math-grpo-adapter API & Inference Endpoint

Training

Base: Qwen/Qwen3.6-35B-A3B (Apache-2.0)
Data: 535 eligible unique items, 534 train and one evaluation item
Procedure: LoRA SFT followed by matched GRPO
LoRA: rank 16, alpha 32
Hardware: four NVIDIA H200 GPUs for the recorded SFT variant
GRPO: 60 logged steps, 3,629.4258 seconds

Raw training questions are not redistributed by this repository.

Evaluation

Table with columns: Stage, Hard128, MMLU-Pro64, MuSR64, BBEH61
Stage	Hard128	MMLU-Pro64	MuSR64	BBEH61
Base	99	60	39	33
SFT	103	59	44	36
GRPO60	106	61	45	41

BBEH61 used a documented 64k-thinking policy and one seed. Base-to-GRPO McNemar p=0.3018; this is not statistically conclusive evidence of broad reasoning improvement.

Reproducibility and boundaries

Evidence repository v1.1.0
Git commit: 0937ef30ea1e3f4b85ec87d1696b22191b62ec14
The sft-high-v2 revision contains the precursor SFT adapter.
The grpo60-high-v2 tag identifies the audited final adapter.

Use is subject to the base model license. Results establish bounded task specialization, not universal mathematical reasoning gains.

Training

Base: Qwen/Qwen3.6-35B-A3B (Apache-2.0)
Data: 535 eligible unique items, 534 train and one evaluation item
Procedure: LoRA SFT followed by matched GRPO
LoRA: rank 16, alpha 32
Hardware: four NVIDIA H200 GPUs for the recorded SFT variant
GRPO: 60 logged steps, 3,629.4258 seconds

Raw training questions are not redistributed by this repository.

Evaluation

Table with columns: Stage, Hard128, MMLU-Pro64, MuSR64, BBEH61
Stage	Hard128	MMLU-Pro64	MuSR64	BBEH61
Base	99	60	39	33
SFT	103	59	44	36
GRPO60	106	61	45	41

BBEH61 used a documented 64k-thinking policy and one seed. Base-to-GRPO McNemar p=0.3018; this is not statistically conclusive evidence of broad reasoning improvement.

Reproducibility and boundaries

Evidence repository v1.1.0
Git commit: 0937ef30ea1e3f4b85ec87d1696b22191b62ec14
The sft-high-v2 revision contains the precursor SFT adapter.
The grpo60-high-v2 tag identifies the audited final adapter.

Use is subject to the base model license. Results establish bounded task specialization, not universal mathematical reasoning gains.

qwen3.6-35b-a3b-math-grpo-adapter

README

Training

Evaluation

Reproducibility and boundaries

Explore FriendliAI today

README

Training

Evaluation

Reproducibility and boundaries