Spakie/DeepSeek-V4-Pro-distill-V2

DeepSeek-V4-Pro-distill-V2

39,830 general-purpose chat and instruction-following examples distilled from DeepSeek-V4-Pro, fact-checked and patched using GPT-5.5 Thinking.

Pipeline

1. 1.Distillation — responses generated via DeepSeek-V4-Pro API
2. 2.Fact-checking — GPT-5.5 Thinking with web search inside Codex reviewed all examples for factual errors, hallucinations and syntax/runtime erros in code examples

Format

Standard chat format, compatible with most SFT frameworks. Each row is one JSON object with a messages array. The dataset includes both single-turn examples and multi-turn conversations:

{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}

Loading

from datasets import load_dataset

dataset = load_dataset("Spakie/DeepSeek-V4-Pro-distill-V2")
train = dataset["train"]

Dataset Structure

• —messages: list of chat messages
• —messages[].role: message role, typically user or assistant
• —messages[].content: message text
• —single-turn rows contain one user/assistant exchange
• —multi-turn rows contain repeated user/assistant exchanges in the same messages array

Coverage

Broad general SFT mix: factual Q&A, coding, creative writing, explanations, translations, rewrites, casual conversation, and multi-turn assistant interactions.

Notes

• —Distilled data — responses reflect DeepSeek-V4-Pro's style and knowledge
• —Use as-is or as a base for further filtering/mixing
• —This does NOT contain reasoning
• —Free to use since I'm not gonna gatekeep distilled data of an open model

Made with ❤️ by Spakie