leonardklin/OpenSciDER-SFT-8K

---
configs:
- config_name: openscider
data_files:
- split: train
path: datafiles/*openscider*.jsonl
- config_name: benchmarks
data_files:
- split: train
path:
- datafiles/airs*.josnl
- datafiles/astrovis*.jsonl
- datafiles/discovery*.jsonl
- datafiles/mle*.jsonl
- datafiles/scicode*.jsonl
- config_name: all
data_files:
- split: train
path: datafiles/*.jsonl
license: apache-2.0
size_categories:
- 1K-10K
task_categories:
- text-generation
tags:
- reasoning
language:
- en
---

<div align="center">
<img src="https://github.com/leonardodalinky/SciDER/blob/a089864587cf1df0d97099f8395505e7ed04caa8/static/images/scider_logo.webp?raw=true" width="400" />
</div>

# OpenSciDER

This is the model repo for OpenSciDER-SFT-8K, a trajectory dataset curated from [SciDER](https://github.com/leonardodalinky/SciDER).

This dataset is collected from the inference trajectories of the Qwen3.6-27B and [OpenSciDER-27B](https://huggingface.co/leonardklin/OpenSciDER-27B) model on DataSciBench, DS-1000, DS-Bench, and ScienceAgentBench. It also contains the benchmark evaluation trajectories on AI-Idea-Bench, AIRS-Bench, AstroVisBench, DiscoveryBench, MLE-Bench, and SciCode.

There are 3 configurations in this dataset:
- `openscider`: the trajectories of Qwen3.6-27B and OpenSciDER-27B on DataSciBench, DS-1000, DS-Bench, and ScienceAgentBench.
- `benchmarks`: the benchmark evaluation trajectories on AI-Idea-Bench, AIRS-Bench, AstroVisBench, DiscoveryBench, MLE-Bench, and SciCode.
- `all`: all the trajectories mentioned above.

You can load the dataset with the HuggingFace `datasets` library:

```python
from datasets import load_dataset

dataset = load_dataset("leonardklin/OpenSciDER-SFT-8K", "openscider")
# or
dataset = load_dataset("leonardklin/OpenSciDER-SFT-8K", "benchmarks")
# or
dataset = load_dataset("leonardklin/OpenSciDER-SFT-8K", "all")
```