open-athena/ultrachat-persona-conversations

UltraChat Persona Conversations

This dataset contains no text copied from its two source datasets. It is built from HuggingFaceH4/ultrachat_200k (train_sft split — real, human-written opening prompts) and nvidia/Nemotron-Personas-USA (synthetic persona profiles), but instead of redistributing their text, every row here carries only the foreign keys — prompt_id and persona_uuid — back to the original row in each source. An enrich.py script is included to reconstruct the full conversation locally by joining against those two datasets yourself, under their own respective licenses.

Everything that is included as text — every conversational turn — was generated by GLM-5.3: it played a persona (conditioned on the sampled Nemotron profile) continuing the real UltraChat opening prompt, and separately played a normal helpful assistant responding to it, in a two-role self-play loop. No real person wrote or received any of these conversations.

How it was built

1. Prompts: the prompt field from all 207,865 rows of ultrachat_200k's train_sft split — real user-written requests (how-tos, writing tasks, analysis questions, etc.).
2. Personas: one row sampled (with replacement, seeded per prompt for reproducibility) from the first shard of Nemotron-Personas-USA (90,910 synthetic USA personas) for each prompt.
3. Conversations: GLM-5.3 played both sides of a chat, seeded with the real prompt as the opening user turn:
   • User side: the sampled persona, continuing the conversation in a voice consistent with its profile (interests, background, register). Each turn it emits a structured {"message", "done"} decision; done: true ends the conversation in character — satisfied, done engaging, or just moving on. Agreement/satisfaction is never forced.
   • Assistant side: a normal helpful assistant, responding naturally to whatever the persona said.
   • A soft cap of 20 total messages forces a stop if a conversation runs long (0.065% of conversations hit it — see Stats).
4. Redaction: the opening prompt's text (messages[0].content) was stripped to null before release, since it is ultrachat_200k's content, not this dataset's. No persona text was ever stored in a conversation record in the first place — persona conditioning only used persona_uuid plus profile fields at generation time.
5. Scope: this covers 200,389 of the 207,865 train_sft prompts (96.4%) — the run was stopped before covering the remaining prompts. prompt_id values not present here are simply prompts that weren't processed; there's nothing wrong with them.

Fields

field	type	description
prompt_id	string	Foreign key into HuggingFaceH4/ultrachat_200k (train_sft split, prompt_id column). Use enrich.py (or your own join) to recover the actual opening prompt text.
persona_uuid	string	Foreign key into nvidia/Nemotron-Personas-USA (train split, shard 0, uuid column). Use enrich.py to recover the persona description and demographics.
messages	list[{role, content}]	The full transcript. messages[0].content is null (redacted — see above); every other turn is GLM-5.3's own generated text.
turns	int	Number of messages in messages
end_reason	string	resolved (persona chose to end it in character) or max_turns (hit the 20-message soft cap)

Using enrich.py

pip install pandas pyarrow
python3 enrich.py --in train.parquet --out enriched.parquet

This downloads the train_sft split of ultrachat_200k (730MB) and shard 0 of Nemotron-Personas-USA (244MB) on first run, joins them in, and writes an enriched local parquet with messages[0].content filled in and a persona (plus persona_age/persona_sex/persona_occupation/ persona_city/persona_state) column added. That enriched file is yours to use locally — please don't redistribute it as-is, since it would contain ultrachat_200k/Nemotron-Personas-USA content subject to their own licenses.

Stats

• 200,389 conversations, one per prompt (200,389 unique prompt_ids, 80,810 unique persona_uuids — personas were sampled with replacement)
• Turn count: min 3, median 5, mean 5.1, max 20
• end_reason: resolved 200,258 (99.93%), max_turns 131 (0.07%)

Known limitations

• Synthetic, not human data, aside from the (redacted, foreign-keyed) opening prompts. Both conversational roles past turn 1 are generated by the same model family.
• Persona/prompt mismatch is not filtered. Personas were sampled independently of prompt content, so some pairings are a poor fit (e.g. a retiree persona handed a prompt asking for JavaScript code). In practice GLM-5.3 sometimes has the persona simply disown the prompt in-character ("I think there's been a mix-up...") and redirect — this happens in roughly 0.6% of a related, similarly-built dataset from the same project; not separately measured here but the same dynamic applies.
• end_reason: resolved describes how the conversation ended in character, not whether the persona was "satisfied" in any evaluative sense — that's not a concept this dataset tracks.
• Not all 207,865 train_sft prompts are covered (see "Scope" above).

License

The conversational content generated for this dataset (everything in messages except the redacted null at index 0) is released under the OpenMDW License ("Open Model, Data and Weights", version 1.1 — see LICENSE). This dataset does not grant you any rights to ultrachat_200k or Nemotron-Personas-USA content — obtain those under their own licenses if you use enrich.py.