Datasetpublic

nyu-dice-lab/wildchat50m-rewild-sft-1118773

wildchat50m-rewild-sft-1118773 A supervised fine-tuning (SFT) dataset formed by the union of three sources, each reformatted to a single canonical conversational schema (WildChat's format is the ground-truth). Single train split, 1,118,773 rows. Canonical schema Column Type Description conversation_hash string Per-row identifier conversation list[{role: string, content: string}] The chat turns model string Provenance / generating-model label… See the full description on the dataset page: https://huggingface.co/datasets/nyu-dice-lab/wildchat50m-rewild-sft-1118773.

sourceHugging Faceotherupdated 4mo agoView on Hugging Face

0likes15downloads

Dataset Card

wildchat50m-rewild-sft-1118773

A supervised fine-tuning (SFT) dataset formed by the union of three sources, each reformatted to a single canonical conversational schema (WildChat's format is the ground-truth). Single train split, 1,118,773 rows.

Canonical schema

Column	Type	Description
`conversation_hash`	`string`	Per-row identifier
`conversation`	`list[{role: string, content: string}]`	The chat turns
`model`	`string`	Provenance / generating-model label

Design choice: WildChat's richer per-message struct (which also carries content_token_ids, country, language, timestamp, toxic, redacted, etc.) was projected down to the SFT-essential `{role, content}` fields so that all three sources share byte-identical Features (required for concatenate_datasets). This yields one clean, valid SFT dataset.

Sources and per-source row contributions

Source	Rows	`model` label
`nyu-dice-lab/allenai_WildChat-1M-Full-Qwen_Qwen2.5-72B-Instruct` (split `train`)	983,073	`Qwen/Qwen2.5-72B-Instruct`
`penfever/tulu-3-tulu_v3.9_personahub_math_interm_algebra_20k` (split `train`)	20,000	`ai2-adapt-dev/tulu_v3.9_personahub_math_interm_algebra_20k` (the `source` value)
`penfever/mmlu-it` (all splits: auxiliary_train + dev + test + validation)	115,700	`mmlu-it`
Total	1,118,773

Reformatting per source

—WildChat: conversation_hash kept; each conversation message projected to {role, content}; model kept (Qwen/Qwen2.5-72B-Instruct).
—tulu: conversation_hash = the row id; conversation = messages projected to {role, content}; model = the row source value.
—mmlu-it: each raw multiple-choice row was templated into a 2-turn conversation. category is the list of answer choices; answer is the (string) integer index of the correct choice. The user turn is the question followed by the choices lettered A. ... / B. ... / ... (variable choice count supported); the assistant turn is "{LETTER}. {choice}" for the correct choice. conversation_hash = SHA-256 of the question; model = mmlu-it. Field semantics were verified on sampled rows of each split before bulk processing; malformed rows (non-integer / out-of-range answer index) were skipped (count reflected in the totals above).

mmlu-it MC -> conversation example

Before (raw row):

question: Find the degree for the given field extension Q(sqrt(2), sqrt(3), sqrt(18)) over Q.
category (choices): ["0", "4", "2", "6"]
answer (index): "1"

After (templated conversation):

user:
Find the degree for the given field extension Q(sqrt(2), sqrt(3), sqrt(18)) over Q.
A. 0
B. 4
C. 2
D. 6

assistant:
B. 4

CONTAMINATION FLAG (read before evaluating on MMLU)

This dataset includes ALL mmlu-it splits as SFT training data, including the `test`, `validation`, and `dev` splits of MMLU. Training on this data therefore exposes the model to MMLU held-out questions and answers.

Anyone who later evaluates a model trained on this dataset against MMLU (or MMLU-IT) must treat those scores as contaminated. If you need a contamination-free variant, it is trivial to rebuild using only the `auxiliary_train` split of penfever/mmlu-it (~99.8K rows) and dropping test/validation/dev.

Provenance

Built on JSC Jupiter via datasets.concatenate_datasets after casting all three sources to identical Features. Public dataset under the nyu-dice-lab org.