voidful/awesome-chatgpt-dataset is ranked 13,999 among these 8 peers; csarron/awesome-emdl is ranked 14,017.

awesome-chatgpt-dataset Unlock the Power of LLM: Explore These Datasets to Train Your Own ChatGPT! Select your own mixed dataset pick whatever dataset you want to use, then merge and upload: Dataset Detail *Sorted by dataset size (small → large). | mit LIMA | 1K | English | LIMA: Less Is More for Alignment. 7K | English | Safety training mixture with vanilla/adversarial prompts and multi-annotator labels. | odc-by Berkeley Function Calling Leaderboard (BFCL) | 2K | English + code | Function/tool-calling eval covering parallel/multi

-call scenarios across languages. | - im-feeling-curious | 3K | English | Extract from Google’s “I’m Feeling Curious” facts. | - Puffin | 3K | English | Exactly 3,000 multi-turn examples; each response via GPT‑4. 0 cc_sbu_align | 4K | English | MiniGPT‑4 alignment data (image–text). | bsd-3-clause QA-Feedback | 4K | English | Re‑constructed ASQA with human feedback. | - SLF5K | 5K | English | Summarization with Language Feedback (5K unique samples). 0 blended_skill_talk | 7K | English | 7k conversations blending personality, empathy, and knowledge

. | - GSM‑IC | 8K | English | Grade‑School Math with Irrelevant Context (distractor sentences). | - ChatAlpaca‑10K | 10K | English | 10,000 multi‑turn conversations (Alpaca‑based). 0 PKU‑SafeRLHF‑10K | 10K | English | First‑round Safe‑RLHF data with safety preferences. | - Dolly‑15K | 15K | English | 15k instruction records crowdsourced by Databricks. 0 WebGPT (comparisons) | 20K | English | Human preference comparisons for WebGPT reward modeling. | - CodeAlpaca‑20K | 20K | English | 20,022 instruction–code pairs for code generation. | - HelpSteer2 | 21K | English | Open-source helpfulness data for reward

models and preference learning. 0 openapi-function-invocations‑25k | 25K | English | Synthetic + extracted OpenAPI function-call traces. | mit LongForm | 28K | English | Reverse‑instruction long‑text generation dataset. | mit Chatbot Arena Conversations | 33K | English | 33K cleaned Arena chats with pairwise preferences. | - HC3 | 37K | English, Chinese | 37,175 instructions with human vs LLM answers. | - Anthropic HH Golden | 45K | English | Helpful & Harmless preference data; golden subset. | - Mol‑Instructions | 48K | English | Biomolecular instruction dataset for LLMs. 0 RefGPT | 50K | English, Chinese | Cost‑effective pipeline

to generate multi‑turn Q&A with references. | - arxiv‑math‑instruct‑50k | 50K | English | QA pairs derived from arXiv math abstracts. | - arxiv‑math‑instruct‑50k (ArtifactAI) | 51K | English | T5‑generated questions; GPT‑3. | - Traditional Chinese Alpaca | 52K | Traditional Chinese | Alpaca translated by ChatGPT API. 0 Cabrita Dataset | 52K | Portuguese | Alpaca translated to Portuguese. | - Japanese Alpaca | 52K | Japanese | Alpaca translated by ChatGPT API. 0; OpenAI terms Alpaca Dataset | 52K | English | 175 seed instructions completed by OpenAI. 0; OpenAI terms Alpaca Data

Cleaned | 52K | English | Cleaned Alpaca 52K. | - Alpaca GPT‑4 Data | 52K | English | Same prompts, GPT‑4 completions. | - Alpaca GPT‑4 Chinese | 52K | Chinese | GPT‑4 completions for Chinese prompts. | - xLAM Function Calling 60K | 60K | English | Structured tool-calling data for executable agents. 0 Dynosaur | 66K | English | Dynamic growth paradigm for instruction curation. 0 Finance | 69K | English | 68,912 finance‑related instructions. | - WizardLM evol | 70K | English | Evolutionary instruction tuning data (WizardLM). | - Vicuna Dataset | 75K | English | ~100k ShareGPT conversations (curated). | - InstructionTranslation | 80K

| Multi-lingual | M2M‑12B translated instructions (≤512 tokens). | mit Self‑Instruct | 82K | English | 52K seed instructions; 82K I/O pairs.