yen-av/tunix-stem-sft

---
dataset_info:
- config_name: default
features:
- name: prompt
dtype: string
- name: reasoning
dtype: string
- name: answer
dtype: string
- name: domain
dtype: string
- name: source
dtype: string
- name: reasoning_length
dtype: int64
- name: task_id
dtype: int32
- name: text
dtype: string
- name: code
dtype: string
- name: test_list
list: string
- name: test_setup_code
dtype: string
- name: challenge_test_list
list: string
splits:
- name: train
num_bytes: 485450485.0
num_examples: 429645
download_size: 251905581
dataset_size: 485450485.0
- config_name: gsm-textbook
features:
- name: prompt
dtype: string
- name: reasoning
dtype: string
- name: answer
dtype: string
- name: domain
dtype: string
- name: source
dtype: string
- name: reasoning_length
dtype: int64
splits:
- name: train
num_bytes: 472726095.0
num_examples: 428681
download_size: 250545416
dataset_size: 472726095.0
- config_name: mbpp
features:
- name: task_id
dtype: int32
- name: text
dtype: string
- name: code
dtype: string
- name: test_list
list: string
- name: test_setup_code
dtype: string
- name: challenge_test_list
list: string
- name: prompt
dtype: string
- name: reasoning
dtype: string
- name: answer
dtype: string
- name: domain
dtype: string
- name: source
dtype: string
- name: reasoning_length
dtype: int64
splits:
- name: train
num_bytes: 2114530
num_examples: 964
download_size: 1043277
dataset_size: 2114530
- config_name: sft-20k
features:
- name: prompt
dtype: string
- name: reasoning
dtype: string
- name: answer
dtype: string
- name: domain
dtype: string
- name: source
dtype: string
- name: reasoning_length
dtype: int64
- name: task_id
dtype: int32
- name: text
dtype: string
- name: code
dtype: string
- name: test_list
list: string
- name: test_setup_code
dtype: string
- name: challenge_test_list
list: string
splits:
- name: train
num_bytes: 22559496.812348995
num_examples: 20000
download_size: 12683986
dataset_size: 22559496.812348995
- config_name: verifiable
features:
- name: prompt
dtype: string
- name: reasoning
dtype: string
- name: answer
dtype: string
- name: domain
dtype: string
- name: source
dtype: string
- name: reasoning_length
dtype: int64
- name: answer_type
dtype: string
- name: task_id
dtype: int32
- name: text
dtype: string
- name: code
dtype: string
- name: test_list
list: string
- name: test_setup_code
dtype: string
- name: challenge_test_list
list: string
splits:
- name: train
num_bytes: 213820869
num_examples: 211163
download_size: 110783444
dataset_size: 213820869
- config_name: verifiable-20k
features:
- name: prompt
dtype: string
- name: reasoning
dtype: string
- name: answer
dtype: string
- name: domain
dtype: string
- name: source
dtype: string
- name: reasoning_length
dtype: int64
- name: answer_type
dtype: string
- name: test_list
list: string
- name: test_setup_code
dtype: string
- name: challenge_test_list
list: string
splits:
- name: train
num_bytes: 19923964.11431358
num_examples: 20000
download_size: 8961900
dataset_size: 19923964.11431358
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- config_name: gsm-textbook
data_files:
- split: train
path: gsm-textbook/train-*
- config_name: mbpp
data_files:
- split: train
path: mbpp/train-*
- config_name: sft-20k
data_files:
- split: train
path: sft-20k/train-*
- config_name: verifiable
data_files:
- split: train
path: verifiable/train-*
- config_name: verifiable-20k
data_files:
- split: train
path: verifiable-20k/train-*
---

# Reasoning Training Dataset for Tunix Competition

Reasoning dataset for training 1-2B thinking models on math, coding, and science problems.

## Sources
- **GSM8K**: Grade school math with human reasoning traces
- **TextbookReasoning**: STEM problems with step-by-step solutions
- **MBPP**: Basic Python Programming prompts, with reasoning traces generated by gpt-oss-20b

## Format
Each example contains:
- `prompt`: The problem statement
- `reasoning`: Step-by-step reasoning
- `answer`: Final answer
- `domain`: 'math'|'physics'|'cs'|'chemistry'|'biology' | 'code'
- `source`: Original dataset name

## Intended Use
Train with Tunix SFT for Gemma to learn concise reasoning traces matching competition format:
```
<reasoning>step-by-step thinking</reasoning>
<answer>final answer</answer>
```

## License
Inherits licenses from source datasets