NovaSky-AI

NovaSky-AI/SkyRL

SkyRL: A Modular Full-stack RL Library for LLMs

20 good first / help-wanted issues · Python · last activity Sep 26, 2026

2.3K stars 439 forks 2.3K watchers Python Apache License 2.0

View on GitHub Website

20 Open Issues Need Help Last updated: Sep 26, 2026

Open Issues Need Help

View All on GitHub

[skyrl-train] Fix verifiers integration about 1 month ago

bug help wanted skyrl-train

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

Add delete checkpoint endpoint to prevent Tinker checkpoint disk exhaustion 4 months ago

good first issue

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

[train] Migrate FlashRL fork to use the latest vLLM version 6 months ago

bug help wanted

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

[train] Support streming mini-batch (i.e. mini-batch-level overlapping async training) 6 months ago

good first issue skyrl-train

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

[train][FullyAsync] Support customizable weight sync frequency 8 months ago

good first issue skyrl-train

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

[train][StepWise] Add auto prefix-aware merging in step wise training 8 months ago

good first issue skyrl-train

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

[Harbor] Support agents beyond terminus 2 such as open hands 8 months ago

good first issue

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

[tx] Implement efficient kernel for ragged_dot that supports expert parallelism 10 months ago

good first issue tx

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

[skyrl-train] Track inference vs trainer logprobs mismatch as a logger metric 11 months ago

good first issue skyrl-train

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

[skyrl-train][logging] Add metrics from all nodes using tracker (i.e. when using WandB) 11 months ago

good first issue skyrl-train

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

[cleanup] Pretty print rollout logging 11 months ago

good first issue skyrl-train

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

[skyrl-train] flash-rl installation broken 11 months ago

bug help wanted skyrl-train

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

[tx] Reactivate type checking with `ty` 11 months ago

good first issue tx

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

[tx] Implement top_k and top_p and eos / stop tokens for sampling 12 months ago

good first issue tx

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

Checkpointing Improvements 12 months ago

help wanted skyrl-train

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

[tx] Introduce proper types for the return values of Qwen3Model.__call__ and Qwen3ForCausalLM.__call__ about 1 year ago

good first issue tx

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

Implement more LoRA layers about 1 year ago

good first issue tx

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

Allow storing model checkpoints on cloud storage about 1 year ago

good first issue tx

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

Rename tx/tinker/models.py to tx/tinker/db_models.py about 1 year ago

good first issue tx

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python

Implement properly downloading the checkpoints via the server about 1 year ago

good first issue tx

NovaSky-AI/SkyRL

2.3K

SkyRL: A Modular Full-stack RL Library for LLMs

Python