Most dialog SFT datasets teach models to plan like an external narrator; this collection instead trains the model to think and respond from inside a character’s head. That shift matters when you want a chatbot that maintains persona, motivations, and in-character reasoning across long roleplays rather than intermittently narrating its strategy.
What Sets It Apart
- Character‑POV CoT traces: Assistant messages begin with
<think>...</think>inner monologues that read as the character’s private reasoning, not an external plan. This preserves voice, memory, and attitude in downstream generations. - Distilled teacher provenance: The dataset was distilled from a larger Gryphe/Aesir source using a high-capacity teacher (deepseek-v4-pro) with an explicit review pipeline; each sample records model/distillation metadata so you can filter by provenance or quality.
- Practical filters & review flags: Samples are labeled keep/borderline and include reviewer_verdict metadata plus safety filtering (mixed SFW/NSFW; some adult RP removed). Authors provide guidance on dropping borderline examples for strict training.
- Compact & ready: ~14,349 assistant turns across ~1,973 conversations in parquet/optimized-parquet format, compatible with Hugging Face datasets, pandas, and polars for quick loading.
Who it's for — and tradeoffs
Great fit if you are fine-tuning a conversational model to maintain persona and inner-monologue style reasoning (SFT for character AI, creative narrative agents, or RP-focused assistants). Look elsewhere if you need strictly SFW corpora, large-scale pretraining data (this is 1K–10K samples), or if your deployment prohibits using distilled traces derived from third-party closed teacher models.
Where it fits
Use as a targeted SFT subset to improve persona persistence and character-driven decision-making in chat agents. Combine with broader conversational datasets for coverage; use the reviewer_verdict filter to create a strict training split.
Notes: license inherits the source (verify before commercial use). The dataset page contains community discussions, further quality notes, and a citation entry for reproducible research.
