---
license: apache-2.0
base_model: Qwen/Qwen3.5-4B
tags:
- reinforcement-learning
- grpo
- trl
- harbor
- data-agent
- agentic
pipeline_tag: text-generation
library_name: transformers
---

# data-agent-4b-normal-best (v0)

A **4B** data-science agent finetuned from [`Qwen/Qwen3.5-4B`](https://huggingface.co/Qwen/Qwen3.5-4B) with
**GRPO** (online RL) to solve data-analysis tasks in a sandboxed bash environment. This repo holds
the **best-eval** checkpoint (peak pass@4, step 200) of the `4b-normal` run.

## Training
- **Method:** GRPO (Group Relative Policy Optimization) via [TRL](https://github.com/huggingface/trl).
- **Environment:** [Harbor](https://github.com/huggingface/trl) task spec + **E2B** cloud sandboxes; single `bash` tool, answer submitted to `/workdir/answer.txt`.
- **Dataset:** [`AdithyaSK/data_agent_rl_environment_train`](https://huggingface.co/datasets/AdithyaSK/data_agent_rl_environment_train).
- **Schedule:** 1 epoch (1119 steps), 8 generations/prompt, KL-anchored to the reference. Tasks were presented in random order (standard).
- **This checkpoint:** step **200** (best-eval).

## Evaluation
Agentic pass@k on the held-out [`data_agent_rl_environment_eval`](https://huggingface.co/datasets/AdithyaSK/data_agent_rl_environment_eval) suite (366 tasks, 4 samples/task, unbiased estimator):

| metric | base (Qwen3.5-4B) | this model | Δ |
|---|---|---|---|
| pass@1 | 0.600 | **0.660** | +0.060 |
| pass@2 | 0.677 | **0.745** | +0.068 |
| pass@3 | 0.715 | **0.779** | +0.064 |
| pass@4 | 0.740 | **0.802** | +0.062 |

## Usage
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("AdithyaSK/data-agent-4b-normal-best", revision="v0", torch_dtype="bfloat16")
tok = AutoTokenizer.from_pretrained("AdithyaSK/data-agent-4b-normal-best", revision="v0")
```

*Part of the **data-agent v0** release. Served non-thinking with a single `bash` tool (Qwen tool-calling).*