Release v0.1.10

What's Changed

• fix(geo3k): avoid duplicate turn-end tokens by @k21993 in #131
• fix(metrics): report the IS filter fraction per sequence, not per token by @k21993 in #133
• fix(critic): scale the MoE aux loss by the critic's own window by @k21993 in #134
• fix(recipe): resubmit the renamed GLM-5.2 Slurm script by @MaxFreedomPollard in #135
• fix(sft): keep VLM samples within max_len by @k21993 in #137
• docs: polish README wording and fix stale/incorrect details by @hijkzzz in #138
• fix(eval): count rollouts in eval_num_samples by @k21993 in #140
• fix(geo3k): stop on nested boxed answers by @k21993 in #139
• fix(packaging): publish an architecture-independent wheel by @k21993 in #143
• docs(vllm): correct stale KV-cache claims around weight sync by @sysuTobo in #141
• fix(geo3k): mark turn-limit exits as truncated by @k21993 in #142
• docs(recipes): align MoE dispatcher comments with defaults by @k21993 in #147
• feat(lora): LoRA for SFT and RL via AutoModel-native PEFT by @sysuTobo in #144
• fix(agents): preserve chat rollout truncation by @k21993 in #145
• refactor(lora): build PeftConfig once in BaseModel; trim comments; te… by @hijkzzz in #148
• fix(agents): skip the tool on a truncated final turn; document the Ch… by @hijkzzz in #149
• Chore/bump vllm 0.30 automodel by @hijkzzz in #150
• Fix/review followups by @hijkzzz in #151
• ci: on-demand two-GPU asynchronous RL end-to-end workflow by @hijkzzz in #152
• ci(e2e-rl): run on a PR via the e2e-rl-test label; post every metric'… by @hijkzzz in #153

New Contributors

• @MaxFreedomPollard made their first contribution in #135
• @sysuTobo made their first contribution in #141

Full Changelog: v0.1.9...v0.1.10