GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
zhongweixie
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
zhongweixie
Model Tree
Qwen/Qwen3-VL-8B-Instruct
Input Modalities
Output Modalities
Supported Functionality
| 项 | 值 |
|---|---|
| 实验名 | exp_rl_v5_v13_gdpo_dim_scoring_stage3 |
| 算法 | GDPO |
| 变体 | 启用 rubric overlay + 分维度打分 (CLAW_RL_DIM_SCORING=1，env 换为 claw_rl.claw_env_dim_scoring，reward 后处理换为 claw_rl.group_filter_dim)。 |
| 起点 | stage3 SFT checkpoint (exp_stage3_v2_merged)，其上游为 Qwen3-VL-8B-Instruct |
| 最终 checkpoint | iter_0000560 |
| 训练量 | 561 个 rollout step，约 3 个 epoch，187 条 prompt |
| samples / prompt | 8 |
| rollout batch | 1 |
| global batch | 8 |
| 硬件 | 8xH800，TP=8，colocate |
| 奖励 judge | qwen/qwen-2.5-72b-instruct (OpenRouter) |
学习率、KL 系数等由 slime 的 formal profile 提供，未在启动脚本中显式设置，
故此处不列出，以免给出未经核实的数值。
目前没有可信的 held-out 评测数字。 不要把本模型当作已验证优于其起点。
原因：
ninja 未能启动，缺少可比基线。起点 SFT checkpoint 记录的 eval=0.506 是起点的分数，不是本模型的分数。
python
from transformers import Qwen3VLForConditionalGeneration, AutoProcessormodel = Qwen3VLForConditionalGeneration.from_pretrained("zhongweixie/qwen3vl-8b-claw-rl-gdpo-dim-scoring-stage3-iter560", dtype="auto", device_map="auto")processor = AutoProcessor.from_pretrained("zhongweixie/qwen3vl-8b-claw-rl-gdpo-dim-scoring-stage3-iter560")
由 slime/tools/convert_torch_dist_to_hf.py 从 Megatron 分布式 checkpoint
/aifs4su/hansirui_3rd/rl-checkpoints/exp_rl_v5_v13_gdpo_dim_scoring_stage3/iter_0000560 转换而来，缺失权重与 tokenizer/config 取自
exp_stage3_v2_merged_sharded(即该次 RL 训练实际使用的基座)。