Note The article: The ultimate guide to multi-harness RL

Note Start here: solve a data task in the SETA whitebox playground. These same Bash tools are used by GRPO.

Note Deprecated native OpenCode interface, kept for the historical comparison. Use Harbor with harness="opencode" for new runs.

Note RL training dashboard: LFM and Qwen reward curves, per-harness checkpoint pass@1, tool/token efficiency and evaluation coverage. Opens the LFM comparison by default.

Note Harbor multi-harness blackbox. Browse tasks, connect a model, run a harness and inspect live traces.

Note SFT training dashboard: training loss, token accuracy and checkpoint evaluations. Use the LFM OpenCode and multi-harness runs for the article’s SFT comparison.

LiquidAI/LFM2.5-2.6B

Text Generation • 3B • Updated Aug 24 • 104k • 799

Note Base model for the main RL and SFT runs

Qwen/Qwen3.5-2B

Image-Text-to-Text • 2B • Updated Mar 2 • 4.75M • • 426

Note Base model for the earlier Qwen runs

Note SmolDataEnvs training tasks in Harbor format

Note SmolDataEnvs held-out test tasks (250) in Harbor format

Note SmolDataEnvs source tasks

Note Multi-harness SFT data from Qwen3.8-27B rollouts, per harness and pre-tokenized

Note Raw successful teacher rollouts across four harnesses

Note Browse the Harbor tasks file by file

FineEnvs/LFM2.5-2.6B-multiharness-RL

Text Generation • 3B • Updated 6 days ago • 327 • 2

Note TRL Async GRPO; main: step 1,000, 54.2% pass@1 across four harnesses. Best: [step-700](https://huggingface.co/FineEnvs/LFM2.5-2.6B-multiharness-RL/tree/step-700), 54.6%.

FineEnvs/LFM2.5-2.6B-opencode-RL

Text Generation • 3B • Updated 6 days ago • 281

Note TRL Async GRPO; main: step 1,000, 52.3% pass@1 across four harnesses. Best: [step-900](https://huggingface.co/FineEnvs/LFM2.5-2.6B-opencode-RL/tree/step-900), 52.4%.

FineEnvs/LFM2.5-2.6B-multiharness-SFT

Text Generation • 3B • Updated 6 days ago • 260

Note TRL SFT; main: step 4,484, 43.1% pass@1 across four harnesses.

FineEnvs/LFM2.5-2.6B-opencode-SFT

Text Generation • 3B • Updated 6 days ago • 292 • 1

Note TRL SFT; main: step 1,208, 47.5% pass@1 across four harnesses.

FineEnvs/Qwen3.5-2B-multiharness-RL

Text Generation • 2B • Updated 6 days ago • 295 • 1

Note TRL Async GRPO; main: step 500, 37.0% pass@1 across four harnesses.

FineEnvs/Qwen3.5-2B-opencode-RL

Text Generation • 2B • Updated 6 days ago • 261

Note TRL Async GRPO; main: step 700, 39.5% pass@1 across four harnesses.

FineEnvs/Qwen3.5-2B-opencode-standalone-RL

Text Generation • 2B • Updated 6 days ago • 261

Note TRL Async GRPO; main: step 1,000, 29.8% pass@1 across four harnesses.