Skip to content

Dataset · r0b0tlab

DeepSeek V4 Pro reasoning traces for agents

r0b0tlab/deepseek-hermes-reasoning-traces

19,331 multi-step reasoning dialogues in ChatML + Hermes format in English for training local agent models.

What's inside

Contains 19,331 multi-step reasoning dialogues generated by the DeepSeek V4 Pro model in English. The data is formatted in ChatML and Hermes formats, split into train (16,431), valid (1,933), and test (967) splits, and divided by token volume for different VRAM sizes (nano, budget, standard). The dataset is designed for fine-tuning local models to act as Hermes Agent instances.

What you can do with it

  • LoRA fine-tuning of local LLMs to act as autonomous agents (Hermes Agent)
  • Training smaller models in step-by-step reasoning under limited VRAM
  • Testing the quality of agent reasoning chains on a held-out test split
Tasks
Text generation
Languages
English
Size
10K–100K rows
Downloads this month
293
Likes
42

How to load it

from datasets import load_dataset

ds = load_dataset("r0b0tlab/deepseek-hermes-reasoning-traces")