GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
arvindcr4
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
arvindcr4
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
platform_hybrid/experiments/tinker-runs/results/h2h2_grpo_Qwen3.5-4B_s123_G8_Dmedium_R16_S15.json; Tinker run fae28299-45a5-53d9-8d09-34d4e9edacf0:train:0,
checkpoint tinker://fae28299-45a5-53d9-8d09-34d4e9edacf0:train:0/sampler_weights/final.platform_tinker/tinkerrl/grpo_cli.py.Qwen/Qwen3.5-4B; LoRA rank 16, α=32, all linear layers (attn, MLP, unembed).main train split ("medium" difficulty bucket). Reward: binary exact match of the final numeric answer.| Metric | Value |
|---|---|
| Mean reward, first 5 steps | 0.4875 |
| Mean reward, last 10 steps | 0.6875 |
| Peak step reward | 1.00 |
| Steps with zero loss (all groups zero-variance) | 20% |
No held-out evaluation was run for this adapter. Step rewards come from 2 prompts × 8 samples (16 rollouts), so they are very noisy, and no confidence interval is meaningful for a single 15-step run.
The ZVF numbers in the thesis and paper do not come from this adapter. They come from
stored per-prompt reward tensors (Qwen3-8B, G=8, 3 seeds: ZVF 0.130 / 0.190 / 0.155)
and from W&B project-level runs. To first order, ZVF is the binomial identity
mean_i[p_i^G + (1−p_i)^G] over per-prompt solve rates. See REPRODUCE.md, Part B.
A single seed with 15 steps. This is a provenance artefact, not a model for deployment.
Apache-2.0, matching Qwen/Qwen3.5-4B. GSM8K is MIT-licensed.