Executor-Quotient Stage 1 matched PPO artifacts

本 model repo 保存 Stage 1 frozen matched Point-ratio / robust-core Hybrid-ratio PPO kill gate 的四个 trainable-only final states、训练 receipts、test predictions、BrowserGym real-click executions 与机械报告。

结论

机械 verdict 为 NO-GO for robust-core PPO as the paper mechanism。Hybrid 相对 Point 的 combined test success gain 为 +2.17pp,task-seed stratified paired-bootstrap 95% interval 为 [+0.56,+3.78]pp;该 gain 未达到冻结的 +5pp,且 seed 29 Hybrid final approximate KL 比 Point 高 0.00948 > 0.005。

Provenance

  • Git repo:https://github.com/luojiaxuan/executor-quotient-gui-policy
  • Artifact repo:gavinlaw/executor-quotient-gui-policy-stage1-ppo;四个 LFS SHA 与小型 receipts 已在 revision 80b1ce627c9d6d30c3b09d43a6665eb1f4e9a93a 回读验证;
  • 生成代码 commit:1eb555deba9ae26b840fd1b3d2436104e6abd539
  • 初始化模型:gavinlaw/executor-quotient-gui-policy-stage1@bffc7aaac1fa7656854bd27b6be2cdd747c990b2
  • rollout / robust-core artifacts:gavinlaw/executor-quotient-miniwob-ppo-rollouts@699af20ff2351b9d99510bf83977c45431952ef2
  • test states:gavinlaw/executor-quotient-miniwob-event-lattice@5bc2d84d1cf30b374e6922a1191b55b4fa6b44bd
  • 冻结合同:docs/stage1_ppo_matched_contract.md
  • 机械结果:docs/stage1_ppo_result.md 与 results/stage1/ppo-gate-report.json

Layout

  • models/{point,hybrid}-{11,29}.pt:只含 trainable_model 与 global_step 的 final states;
  • artifacts/training/<arm>/:config、optimizer-step log、full-buffer atom KL 与 completion receipt;
  • artifacts/predictions/<arm>/:900 条 disjoint test predictions;
  • artifacts/execution/<arm>/:900 条 BrowserGym real-click results 与 summary;
  • artifacts/report.json:10,000-sample paired-bootstrap mechanical gate report;
  • manifest.json:文件 SHA256、输入 revisions 与 schema。

参数文件通过 code/compact_stage1_checkpoints.py --output-dir ... 从 final optimizer checkpoints 导出。模型结构由 Git 代码定义,加载时读取 trainable_model mapping;这些参数不是独立的 Transformers from_pretrained checkpoint。

许可

Built with Qwen. 这里的参数文件是 Qwen/Qwen2.5-VL-3B-Instruct(revision 66285546d2b821cf421d4f5eb2576359d3770cd3)的衍生作品。基座按 Qwen Research License Agreement 发布,这些参数因此只能用于非商业的研究或评测,商用需要向 Alibaba Cloud 申请许可。LICENSE 是该协议的副本,NOTICE 是协议要求的署名声明。

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading

Model tree for gavinlaw/executor-quotient-gui-policy-stage1-ppo

Finetuned
(939)
this model