Back to Gap Map Model components / Training & synthetic datasets

WildChat-1M

open / Overall score: 3.5

WildChat-1M is a corpus of real user conversations with ChatGPT, drawn from about one million conversations gathered by offering online users free access to GPT-3.5 and GPT-4. It covers 68 detected languages and records timestamps, hashed IP addresses, country and state, the OpenAI model used, and moderation scores. The public release omits conversations flagged as toxic or as containing sensitive information, and is published by the Allen Institute for AI on Hugging Face.

Openness

5 medium confidence

5.0

license: odc-by
gated: false per HF API
dataset_card: present

ODC-BY licensed and reported ungated by the HF API, though Ai2 datasets sometimes show an access notice.

https://huggingface.co/api/datasets/allenai/WildChat-1M recorded 2026-09-29
The Hub API record for allenai/WildChat-1M carries the tags "license:odc-by" and "format:parquet", a gated flag of false, and a description opening "Dataset Card for WildChat".
https://huggingface.co/datasets/allenai/WildChat-1M/raw/main/README.md recorded 2026-09-29
The raw dataset card lists "License:** [ODC-BY](https://opendatacommons.org/licenses/by/1-0/)" with sections "Data Fields" and "Licensing Information", and notes the toxic superset WildChat-1M-Full "requires approval and justification for why toxic data is needed".

Adoption

3 high confidence

3.0

Adoption is measured by Hugging Face downloads of the dataset over the trailing month, the only usage channel on record for WildChat-1M.

https://huggingface.co/api/datasets/allenai/WildChat-1M recorded 2026-08-13
19,540 downloads in the trailing 30 days for allenai/WildChat-1M

Capability

4 medium confidence

4.0

A sustained cross-recipe SFT ingredient (Tulu 3 2024, OLMo 3 Dolci 2025), but its own ablation is weak/dated and it does not lead comparisons.

https://arxiv.org/html/2405.01470v1 recorded 2026-09-29
The WildChat paper says its own fine-tuned model "outperforms other open-source models of the same size, although it significantly underperforms proprietary models".
https://huggingface.co/datasets/allenai/dolci-instruct-sft/raw/main/README.md recorded 2026-09-29
The Dolci Instruct SFT card, used to train Olmo 3 7B Instruct SFT, lists "WildChat with upgraded responses from GPT-4.1 (ODC-BY-1.0), 302,406 prompts" among its sources.
https://huggingface.co/datasets/allenai/tulu-3-sft-mixture/raw/main/README.md recorded 2026-09-29
The Tulu 3 SFT mixture card lists "WildChat GPT-4](https://huggingface.co/datasets/allenai/WildChat-1M) (ODC-BY-1.0), 100,000 prompts" among its sources.

Verified 2026-09-27