PyRUA-Lean Coder v0.1-4B¶
DAGroup-PKU/PyRUA-Lean-Coder-v0.1-4B
PyRUA-Lean Coder v0.1-4B is Qwen3.5-4B trained with GRPO in the PyRUA-Lean framework, on 45 LIBERO-PRO task instances, as its planner: it writes the cells, the primitives do the rest. On 655 task instances it never saw in training it goes from 18.8% to 44.1%, above the Qwen3.5-27B base model (29.2%).
Results¶
Success rate (%), PyRUA-Lean framework:
| Planner | Trained on (45) | LIBERO-PRO other (155) | RoboTwin 2.0 (250) | RC365 atomic (90) | RC365 composite (160) | All unseen (655) |
|---|---|---|---|---|---|---|
| Qwen3.5-4B | 37.8 | 13.5 | 14.0 | 52.2 | 12.5 | 18.8 |
| Qwen3.5-27B | 64.4 | 38.1 | 20.8 | 65.6 | 13.1 | 29.2 |
| PyRUA-Lean Coder v0.1-4B | 71.1 | 37.4 | 52.4 | 76.7 | 19.4 | 44.1 |
RoboTwin 2.0 and both RoboCasa365 sets are benchmarks it was not trained on at all; the LIBERO-PRO column of the unseen part holds that benchmark's other tasks.
Running the model¶
An open model runs behind any OpenAI-compatible endpoint (a local SGLang or vLLM server, for example) and is driven
by the in-process loop of pyrualean play --runtime openai (Agent Configuration). With SGLang:
python -m sglang.launch_server --model-path DAGroup-PKU/PyRUA-Lean-Coder-v0.1-4B --port 30000 \
--context-length 65536 --tool-call-parser qwen3_coder --reasoning-parser qwen3 \
--constrained-json-disable-any-whitespace
Then one LIBERO-PRO episode, on a task the model never saw in training (RPent's LIBERO-PRO stack installed as the Quick Start describes):
export RPENT_ROOT=/path/to/rpent OPENAI_API_KEY=none
pyrualean play --robot libero --arm cells --suite libero_goal_swap --task 3 --seed 0 \
--runtime openai --base-url http://127.0.0.1:30000/v1 --api-key-env OPENAI_API_KEY \
--model DAGroup-PKU/PyRUA-Lean-Coder-v0.1-4B \
--openai-max-tokens 8192 --openai-tool-choice required --openai-temperature 1.0 \
--openai-extra-body '{"top_p": 0.95, "top_k": 20}' \
--images on-demand --feedback pure --guides --max-turns 40 --budget-s 5400 \
--out runs/coder-demo --cuda-device 0
pyrualean summarize runs/coder-demo
These are the settings of the paper's evaluation: a reply of at most 8,192 tokens, a context of 65,536 tokens (an episode that outgrows it ends there), at most 40 code cells and 90 minutes.