neooriginal/neoagent-chat

NeoAgent Chat

A supervised fine-tuning dataset for NeoAgent (by NeoLabs): an assistant that texts like a sharp, warm friend instead of like a customer-service bot, while staying honest about being an AI.

• 59,121 train / 1,828 validation conversations
• English (86%), German (10%), mixed EN/DE (~4%)
• Built with NeoForge, a curation pipeline that turns real human conversations into training examples

What makes it different

Most chat SFT data is written by an LLM. Here, it isn't:

• User turns are real human messages, kept verbatim (typos, slang, lowercase, emoji and all). The pipeline only selects, trims and merges consecutive messages; it never writes user text. Code checks verify that every user turn reuses at least 90% of the original words.
• Assistant turns are built from real replies in the same thread (another human, a forum volunteer, or an old chatbot). An LLM acts only as an editor: it cuts, merges, fixes factual errors with minimal changes and nudges the tone. Code checks enforce high word reuse from the original reply (≥70% for chat sources, ≥55% for chatbot/forum).
• Human-life claims are removed. Replies where the original speaker talks about their own body, job, location, family or past actions are cut or dropped, so the model doesn't learn to pretend it's a person.

Quality filtering

Each candidate goes through:

1. A keyword prefilter for unsafe and sensitive content
2. LLM curation (editor-only, see above)
3. Code checks: verbatim user text, reply word reuse, contiguous merges, junk/links, banned assistant phrases, emoji only when the user used emoji, length limits for casual messages
4. Two independent LLM judges, keeping the lower score per criterion (standalone, helpfulness, length fit, naturalness, mirroring, personality, persona, coherence), plus hard-fail rules (invented user text, factual errors, unsafe content, claiming to be human, sycophancy, wrong language)
5. One optional revise round when the judges consider a draft fixable

Only around 10% of candidates were accepted.

Format

JSON Lines, one conversation per line:

{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}],
 "meta": {"source": "discord_unveiled", "source_format": "discord", "lang": "en", "kind": "smalltalk",
          "origin": ["edited"], "scores": {"standalone": 5, "helpfulness": 5, "...": 5}, "revisions": 1}}

There is no system prompt; add your own at training time. meta.kind is a free-text label from the curator, mostly task, question, banter, smalltalk, opinion, advice, support or identity. meta.scores are the minimum of both judges on a 1–5 scale.

Average length is about 1.3 user turns per conversation; most examples are single exchanges.

Sources

Source	Train rows	Kind
allenai/WildChat-4.8M	36,204	real users talking to a chatbot
fvdfs41/Discord-Unveiled	10,243	real Discord conversations
mookiezi/Discord-Dialogues	10,017	real Discord conversations
OpenAssistant/oasst2	2,108	volunteer Q&A
NottDeuYTSch (German YouTube comments)	549	real comment threads

WildChat makes up ~61% of rows and leans toward tasks and questions. For a more casual model, downsample it (e.g. by meta.source) and upsample German.

License and intended use

This dataset is derived from the sources above, and their licenses and terms apply. Check each one before commercial use or redistribution. It is intended for research and for training conversational assistants.

Limitations

• Content comes from real internet conversations. It was filtered, but it can still contain rough language, mild hostility, opinions, or factual errors that the judges missed.
• Some rows (~0.2%) are in languages other than English or German; filter on meta.lang if needed.
• The judges are LLMs, and their scores are noisy.