wildchat-4o-warm

A filtered subset of WildChat-4.8M's gpt-4o conversations, curated for one purpose: to replicate the warm gpt-4o tone without removing capabilities from models.

Use it as SFT / style data. It is deliberately shaped toward conversational voice — warmth, pacing, and human register — with coding, maths, science, marketing, medical, admin and translation content removed so the model learns tone rather than recalling stale technical answers.

Contents

field	type	notes
conversation_hash	string	source identifier, for dedupe and provenance
model	string	which 4o snapshot answered
timestamp	timestamp	conversation start
turns	int	number of messages
messages	list<{role, content}>	the conversation, roles user / assistant

17,636 conversations, ≈0.096 G characters, ≈25 M tokens. Mean ≈1,360 tokens. Snapshots: gpt-4o-2024-08-06 9,767 · gpt-4o-2024-05-13 6,996 · gpt-4o-2024-11-20 873.

Composition (broad): creative writing ~31 %, small talk ~18 %, how-to ~15 %, personal & advice ~13 %, correspondence ~12 %, editing ~8 %, residual QA ~3 %.

How it was built

Derived from allenai/WildChat-4.8M, non-mini gpt-4o rows only, English, then 19 exclusion rules applied to the first user message:

• coding and coding-adjacent requests; translation; maths, physics, chemistry, biology, astronomy
• money, prices, business, marketing, résumés, interviews
• jailbreak attempts; health and medical; anything involving minors (matched on every turn)
• terse one-offs (first message <80 chars and conversation <800 chars)
• travel, legal/admin, tech support, social captions, recommendations
• eval scaffolds ("Analyze the following message…"), non-Latin script, pasted code/markup, SEO briefs
• DIY/repair, life admin, device/app usage, leaked dev questions

Survivors must then match a positive allowlist (creative writing, correspondence, editing, personal advice, how-to, small talk, residual QA). From 896,633 non-mini 4o conversations, 17,636 survive — 2.0 %.

Intended use

Fine-tuning small instruction models toward the warm 4o register while preserving capability. The filtering specifically removes the domains where a 2024-era model's answers are stale, so the tone can transfer without dragging in outdated technical content.

Recommended practice:

• Dedupe before training. 17,636 rows contain only 12,999 distinct first-message prefixes — 26 % are near-duplicates from retries and re-crawls. Duplicates cause memorisation, not style transfer.
• Mask prompt tokens; train on assistant turns only.
• Hold out a few dozen conversations for style and capability evaluation; do not evaluate on the same slice you trained on.
• Over-sample the personal & advice block if warmth is the objective — it is only ~13 % of rows and is the most relational material in the set.

Caveats

• Keyword filtering leaks. Rules match English keywords on the first message, so dev questions phrased without them (GTK, heap, Angular), non-English text mislabelled English, and pasted code all appear at low rates. An LLM-judge pass would be a better filter than more regex.
• Composition shares are classifier estimates from samples, not labels.
• Not a capability dataset: it has been stripped of exactly the domains capability evals test.

Licensing and provenance

The underlying collection is AI2's WildChat, released under ODC-BY. That licence covers the collection, not OpenAI's rights in the text: the assistant turns are gpt-4o output. Treat this artifact as research data derived from model output, not as a grant of rights, and check the terms that apply to you before redistributing or using it commercially.

Source: allenai/WildChat-4.8M. Field of use: style fine-tuning experiments.