Repository Issue Activity (beta)

#947 Recommended renderer for GPT-OSS emits malformed tool-call headers on 8.5% of requests

#945 Stream-minibatch / async RL hangs when a trajectory worker crashes

#943 Harbor retry make_envs() leaks a full sandbox group (and drops the previous generation from cleanup)

#940 SL checkpoint under default submit_ahead=1 can snapshot the next batch

#935 SL resume replays the last checkpointed batch, applying a duplicate optim_step on the same data

#936 IFBench eval imports the checker registry from the wrong module, so it always raises ImportError even when ifbench is installed

#937 BFCL function-call matching ignores extra/hallucinated arguments, scoring them as correct

#896 MMLU-Redux per-subject accuracy always reports a single "unknown" bucket

#929 introducing-inkling safari compatibility issue

#907 `dro` loss cannot mask positions: quadratic penalty applies at advantage-0 tokens, `weights` input rejected

#895 extract_boxed picks the first \boxed{} in eval but the last in math_rl

#893 IFEval drops a constraint when a prompt repeats an instruction id

#891 IFEval scores every "less than" constraint as satisfied

#897 IFEval scores every repeat_prompt constraint as satisfied

#898 IFEval english_capital accepts title case as all capital letters

#889 [harbor_rl] Binary test fixtures silently zero out 7 Terminal-Bench tasks

#551 Expose per-run cost via API and CLI

#601 Docs: llms.txt / llms-full.txt Not found

#857 # Add a self-improving coding agent example with Inkling

#796 Could not reproduce on policy distillation on DeepMath

#860 An SFT example's observation should be the prompt its turn is sampled after (nemotron3, qwen3_5, kimi_k26)

#847 Recipe idea: fail-closed claim validators + LLM judges for production agent post-training

#281 suggested fixes and improvements to rl/train.py

#268 tool_use/search example is misaligned with tinker's key idea

#217 Request for Support Image Input on the OpenAI Compitble API