zero-prod-6k

Synthetic multi-turn character dialogue corpus zero-prod-6k from the MC000 character fine-tuning pipeline (data-gen stage). Each record is a partner-led, alternating two-party conversation: a teacher session playing the character (assistant turns, driven by the zero character bible) and a teacher session playing a sampled PersonaHub partner (user turns) take turns, one LLM call per message. Facts below are read from the run's run-config.json, filter/summary.json and the export itself.

Corpus

Conversations	5,932
Messages	147,530
assistant (character) turns	73,765
user (partner) turns	73,765
assistant words	2,504,295
user words	4,171,329
Turns / conversation	min 2 · median 24 · mean 24.9 · max 40
Opens with	user
Strictly alternating	yes

Record format

One field, messages: a list of {role, content} turns in ChatML role naming, no system prompt. The character speaks as assistant, the partner as user. Every conversation opens with a partner turn and alternates strictly.

from datasets import load_dataset
ds = load_dataset("movingcastles/zero-prod-6k", split="train")
ds[0]["messages"][:2]

train.jsonl carries the identical records, one JSON object per line.

Generation

Requested conversations	6,000
Seed	60704
Turn range	10-40
Association-walk steps	4
Bare-opener share	0.15
Teacher temperature	1.0
Teacher max_tokens	2048
Character-simulation runtime	enabled (sampled initial conditions)
Character bible	zero/v2

Teacher pool

Role-aware endpoint pool: each conversation pins one endpoint per side (seeded, quota-weighted); failover stays within the same model.

Endpoint	Model	Roles	Weight
groq-gpt-oss-120b	openai/gpt-oss-120b	partner	2.0
groq-llama-33-70b	llama-3.3-70b-versatile	character	2.0
groq-llama-31-8b	llama-3.1-8b-instant	character	2.0
groq-gpt-oss-20b	openai/gpt-oss-20b	partner	2.5
groq-qwen3-32b	qwen/qwen3-32b	character, partner	2.0
anthropic-haiku-45	claude-haiku-4-5-20251001	character, partner	0.5
openai-gpt54-mini	gpt-5.4-mini-2026-03-17	character	3.0
openai-gpt54-nano	gpt-5.4-nano-2026-03-17	partner	3.0

Filtering funnel

Generated	5,964 of 6,000 requested (status partial)
Filter input	5,964
Filter rejected (whole conversation)	29
Truncated (conversation cut at first failing turn)	173 conversations, 2,306 messages dropped
Filter passed	5,935
Manual drops	see provenance/filter-MANUAL-DROPS.md
Exported	5,932

Per-rule outcomes

Rules are scoped to the character side (assistant turns). Failed drops the conversation; truncated cuts it at the first failing turn and keeps the prefix.

Rule	Passed	Failed	Truncated
alignment_markers	5,953	2	9
bliss_markers	5,922	1	41
compression	5,959	0	5
conversation_shape	5,964	0	0
embedding_dedup	5,964	0	0
length	5,956	0	8
self_bleu	5,931	3	30
surface	5,857	23	84
length_distribution	5,935	0	0

Manual drops

Records 935, 3816, 5299 removed by hand before export: degenerate PARTNER-side turns from groq-gpt-oss-20b (ellipsis/question-mark litanies; 935 and 5299 also leak a literal <|endoftext|> string). All filter rules are character-scoped, so these passed the rule set. Records remain in generate/ and analyse/ corpora.