datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
tiny-aya-l2-thinker-multilingual-reasoning
Tiny Aya L2 Multilingual Reasoning (44 languages)
Translated multilingual reasoning traces used to train Tiny Aya L2-Thinker.
Each example has the prompt, thinking, and answer in the same non-English language alongside the original texts in English.
Data source
Prompts from AM-DeepSeek-R1-0528-Distilled
Thinking traces and outputs distilled from gpt-oss-120b
Translated with command-a-translate and DeepSeek-V3
Languages (44)
Language
Train… See the full description on the dataset page: https://huggingface.co/datasets/CohereLabs/tiny-aya-l2-thinker-multilingual-reasoning.tiny-aya-l2-thinker-multilingual-reasoning
Tiny Aya L2 Multilingual Reasoning (44 languages)
Translated multilingual reasoning traces used to train Tiny Aya L2-Thinker.
Each example has the prompt, thinking, and answer in the same non-English language alongside the original texts in English.
Languages (44)
Language
Train
Test
Total
Amharic (am)
3,807
448
4,255
Arabic (ar)
22,968
2,538
25,506
Bulgarian (bg)
4,177
452
4,629
Bengali (bn)
3,803
422
4,225
Catalan (ca)
4,251
512
4,763
Czech… See the full description on the dataset page: https://huggingface.co/datasets/erenyeager-1/tiny-aya-l2-thinker-multilingual-reasoning.Thinker-XMLSystem prompt suggestion:
You are a world-class AI system. Always respond in strict XML format with your reasoning steps within the <im_reasoning> XML tag. Each reasoning step should represent one unit of thought. Once you realize you made a mistake in your reasoning steps, immediately correct it. Place your final response outside the XML tag. Adhere to this XML structure without exception.
thinkerA Chain-of-Thought (CoT) dataset that contains traces of complex and sophisticated reasoning, to mimic the "thinking" process of OpenAI's o1. Wrap the contents of the reasoning column in some XML tag (such as <reasoning>).
Raw .jsonl dataset file can be found under the Files and Versions tab.
unfiltered-thinker
Unfiltered-Thinker: A Dataset for Intermediate Cognitive Reasoning
A corpus of 1,909 samples designed to showcase intermediate thinking, cognitive processes, and structured emotional reasoning.
Source: UnfilteredAI/unfiltered-thinker on Hugging Face
⚠️ Content Warning: This dataset contains content that will be considered offensive, disturbing, or explicit. This includes discussions of dark humor, profanity, criminal activity, violence, substance use, and psychological distress. It… See the full description on the dataset page: https://huggingface.co/datasets/UnfilteredAI/unfiltered-thinker.unfiltered-thinker
Unfiltered-Thinker: A Dataset for Intermediate Cognitive Reasoning
A corpus of 1,908 samples designed to showcase intermediate thinking, cognitive processes, and structured emotional reasoning.
Source: UnfilteredAI/unfiltered-thinker on Hugging Face
⚠️ Content Warning: This dataset contains content that will be considered offensive, disturbing, or explicit. This includes discussions of dark humor, profanity, criminal activity, violence, substance use, and psychological distress.… See the full description on the dataset page: https://huggingface.co/datasets/sonic-coder/unfiltered-thinker.nbeerbower-Purpura-DPO-thinker-rawUnfiltered
System promt for creating a dataset:
You are an expert AI assistant specializing in text generation. Your task is to reverse-engineer the thought process that leads to a given textual `response`.
Based on the user's `prompt` and the final `response` text, generate a plausible, detailed reasoning process of an LLM.
This reasoning should cover:
1. **Analysis of the User's Prompt:** Deconstruct the user's request, identifying explicit constraints (like length, format) and implicit… See the full description on the dataset page: https://huggingface.co/datasets/Disya/nbeerbower-Purpura-DPO-thinker-raw.Thinker-XML-2Suggested system prompt:
Respond to each user instruction in an XML format, using <step> tags to document your logical reasoning process step-by-step, while the <output> tag should be reserved for your final communication with the user. Incorporate self-correction by reflecting on prior steps; if a previous thought requires adjustment, add a new <step> to refine your reasoning without altering the original. Include self-reflection by periodically assessing your thought process and noting any… See the full description on the dataset page: https://huggingface.co/datasets/minchyeom/Thinker-XML-2.Thinker-JSONUse for whatever you want.
Made to replicate the thought traces of OpenAI's o1, I'll release RL datasets including DPO soon enough.
For fine-tuning smaller models such as Google's google/gemma-2-2b-it with this dataset, I recommend fine-tuning for 2-3 epochs, the loss will be at around 1.6 at the beginning, and 1.3 by the end of the training job with learning rate of 2e-6.
Suggested system prompt:
Always respond in strict JSON format with a reasoning_steps array and a response field. Each… See the full description on the dataset page: https://huggingface.co/datasets/minchyeom/Thinker-JSON.