Repository Issue Activity (beta)

#920 reset_batch() wipes RejectionSamplingState every batch, so "episode" mode never accumulates across batches

#916 Bug: `reinforce_plus_plus_baseline` silently produces zero advantages when `rollout.n = 1`

#432 context compression

#391 The actor/entropy loss appears somewhat abnormal—it's too large. We launched DeepSWE training based on the Qwen3-4B model, using the following script:

#915 check_correctness / lcb_check_correctness_v2 leave a zombie process after killing a timed-out test run

#914 R1ToolParser.tool_output_end doesn't match DeepSeek's own chat_template tool-output token

#782 Fireworks loss normalization depends on fwd/bwd chunking after #781

#829 Dataset row ids containing ':' silently collapse every rollout into one GRPO group

#901 No train/held-out split for claw_eval and skillsbench

#468 Sandboxed code execution for RL rollouts

#605 AgentWorkflowPPOTrainer: GRPO groups by trajectory.uid instead of prompt/agent key

#897 async_save=true on the default FSDP strategy never writes the checkpoint tracker file, so resume_mode:auto silently restarts from scratch

#816 OpenAIEngine.completion() retokenization fallback can desync completion_ids from logprobs

#655 Dockerfile RUN-replay corrupts Harbor task setup on modal/daytona (line-continuation mangling + double-apply on prebuilt images)

#388 There may be some compatibility conflicts between Transformers 5.0 and older vLLM.

#380 Supporting deep agents and/or automatically generate sub-agents without predefined flow

#354 actor_rollout_ref.actor.loss_agg_mode=seq-mean-token-sum in examples/swe/train_deepswe_32b.sh

#717 rllm train crashes at first validation: 'FixedEvaluatorHooks' object has no attribute 'warm_queue'

#378 Kind Kubernetes settings

#350 ppo_mini_batch_size for Non-Cumulative Agents

#712 rllm sft crashes with FileNotFoundError — SFT backend config YAMLs excluded from built package

#704 How to run PPO for agents?

#355 How to implement trajectory-level masking for "invalid turns" (SimpleTIR) in rllm?

#226 DeepScaleR 8k reproduction collapses after ~300 steps with sequence length explosion

#434 Qwen3.5 support