Results
Solve rate of OpenCode driven by each model on SmolDataEnvs data-analysis tasks, in Hugging Face Sandboxes, scored by each task's verifier. Every task here is built on a Kaggle dataset that no training trace used: the 144 eval tasks and 163 of the 250 test tasks.
Use it in an agent
Serve it with vLLM and point an OpenAI-compatible agent at it. With OpenEnv's Harbor integration:
vllm serve sergiopaniego/Qwen3.5-2B-SmolDataEnvs-SFT-500 --enable-auto-tool-choice --tool-call-parser qwen3_xml \ --default-chat-template-kwargs '{"enable_thinking": false}' openenv harbor rollout --llm-url http://localhost:8000/v1 --model sergiopaniego/Qwen3.5-2B-SmolDataEnvs-SFT-500 \ --dataset FineEnvs/SmolDataEnvs-harbor-eval --task-index 0 --harness opencode --sandbox hf-sandbox \ --reward-key correctness,reward
Training
- Data:
sergiopaniego/smoldataenvs-opencode-traces, config 500: one row per solved task, the
agent's whole conversation (messages) with the tools it was offered (tools). Traces come only from training
tasks whose Kaggle dataset backs no eval task.
- Loss on assistant turns only (
assistant_only_loss=True, TRL's training version of the Qwen3.5 template), in
no-think mode. 3 epochs, learning rate 1e-5 with cosine decay, batch of 8 sequences of up to 16k tokens, one H200.
- Curves: Trackio.
Limitations
- About a quarter of tasks still end in a tool-call loop that hits the 10-minute timeout.
- Trained and evaluated on one task family (data analysis over Kaggle files with OpenCode). Other agents and task
types are untested.