data.md

Dataset Overview

This dataset consists of three main subsets:

Dataset	Description	Path
full	The standard dataset used for all benchmarks in the paper.	data-full
sample	A smaller subset sampled from full for quick testing.	data-sample
visualize	A subset from full used for error visualization.	data-visualize

The data-full directory contains the following subsets:

Category	Subset	Path
Reasoning	code	data-full/code.jsonl
Reasoning	math	data-full/math.jsonl
Knowledge	law	data-full/law.jsonl
Knowledge	medicine	data-full/medicine.jsonl
Knowledge	business	data-full/business.jsonl
Non-English	french	data-full/french.jsonl
Non-English	japanese	data-full/japanese.jsonl
Non-English	chinese	data-full/chinese.jsonl
Skills	translation (en->zh)	data-full/en2zh.jsonl
Skills	translation (zh->en)	data-full/zh2en.jsonl
Skills	summarization	data-full/sum.jsonl
Long-Context	8k ctx	data-full/long8.jsonl
Long-Context	16k ctx	data-full/long16.jsonl
Long-Context	24k ctx	data-full/long24.jsonl

Dataset Construction

This section describes how to build the dataset from scratch.

1. Download the original data from HuggingFace.

   huggingface-cli download --dataset RyokoAI/ShareGPT52K --local-dir share-gpt

   This will download sg_90k_part1.json and sg_90k_part2.json.

2. Modify the DATA_PATH variable in tools/clean.py and tools/clean_for_longctx.py to point to the downloaded JSON files.

3. Register for an API key on the DeepSeek official website and export it as an environment variable.

   export DEEPSEEK_API_KEY="your_api_key"

4. Run the data processing scripts.

   python tools/clean.py
   python tools/clean_for_longctx.py

After the automated construction, we manually deduplicated the data and curated a selection of high-quality question-answer pairs for each task.