---
license: apache-2.0
base_model: Qwen/Qwen3.5-2B
tags:
- reinforcement-learning
- grpo
- trl
- harbor
- data-agent
- agentic
pipeline_tag: text-generation
library_name: transformers
---

# data-agent-2b-curriculum-final (v0)

A **2B** data-science agent finetuned from [`Qwen/Qwen3.5-2B`](https://huggingface.co/Qwen/Qwen3.5-2B) with
**GRPO** (online RL) to solve data-analysis tasks in a sandboxed bash environment. This repo holds
the **final** checkpoint (end of a full 1-epoch run) of the `2b-curriculum` 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 a **difficulty-ranked curriculum** (easy→hard).
- **This checkpoint:** step **1119** (final).

## 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 (Qwen3.5-2B) | this model | Δ |
|---|---|---|---|
| pass@1 | 0.098 | **0.402** | +0.304 |
| pass@2 | 0.168 | **0.509** | +0.341 |
| pass@3 | 0.229 | **0.563** | +0.334 |
| pass@4 | 0.284 | **0.603** | +0.319 |

## Usage
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("AdithyaSK/data-agent-2b-curriculum-final", revision="v0", torch_dtype="bfloat16")
tok = AutoTokenizer.from_pretrained("AdithyaSK/data-agent-2b-curriculum-final", revision="v0")
```

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