Dataset · FineEnvs
Sandbox RL Tasks for Data Analysis and Code
FineEnvs/SmolDataEnvs-harbor-train
Over 5,500 practical data analysis tasks for training small models via RL.
What's inside
The set includes over 5,500 text tasks for reinforcement learning (RL), of which 5,000 are intended for training (broken down by difficulty from simple to complex) and 144 for validation. In each task, an agent is placed in a sandbox with a real dataset and a question requiring data exploration via code. The project focuses on improving the skills of compact models (e.g., 2B parameters) in data science.
What you can do with it
- RL training of compact language models in code and data analysis
- Test curriculum learning (from simple to complex tasks) for agents
- Evaluate LLM ability to independently explore tabular data in a sandbox
- Tasks
- Other
- Size
- 1K–10K rows
- Downloads this month
- 53.9K
- Likes
- 19
Models trained on it
- mradermacher/LFM2.5-2.6B-opencode-RL-i1-GGUF 640
- mradermacher/LFM2.5-2.6B-opencode-RL-GGUF 374
- FineEnvs/LFM2.5-2.6B-multiharness-RLText generation 227
- FineEnvs/LFM2.5-2.6B-opencode-RLText generation 147
- harness-hub/LFM3-1.7B-hermes 114
- harness-hub/LFM3-1.7B-pi 96
- harness-hub/LFM2.5-24B-A2B-pi 94
- harness-hub/LFM2.5-24B-A2B-claude-code 82
How to load it
from datasets import load_dataset
ds = load_dataset("FineEnvs/SmolDataEnvs-harbor-train")