Card contents

Header table — base model, architecture, LoRA spec (r=128, 524M params), training data (60k of 3.66M conversations), quantization (Q4_K_M, 7.4 GB), context (8k), recommended sampling, license.

The persona protocol — documented end-to-end: the exact persona-card fields to give it (CHARACTER / PERSONA / MOTIVATIONS / PERSONALITY / APPEARANCE / MEMORY / STYLE) and the guaranteed 3-part reply structure (<model_think> → <char_think> → screenplay line), plus the [User]: labeling for multi-speaker parsing.

Test performance — both measured benchmarks: - Format adherence (40 held-out prompts, unseen characters): screenplay lines 92.5%→100%, thinking traces 0%→100%, zero user-speech failures, training loss 3.07→0.30 - RoleBench (100 questions, 5 canon characters, LLM-judge): 4.⁴⁄₁₀ in-character, 4.²⁄₁₀ accuracy — per-character table included (Blair Waldorf best at 6.6, Coach Taylor lowest at 3.1)

Strengths (100% protocol reliability, instant original characters, harness-friendly delimiters, no refusals, single-GPU footprint, consistent long-scene voice) and weaknesses stated plainly (weak canon-character trivia — it’s a persona simulator, not a fandom-knowledge model; shallow trace content; protocol needs the card + /api/chat; English-dominant; no safety alignment; Q4 quality cost).

Best use cases ranked (original-character companions first, multi-character engines, IF/DM, game dialogue prototyping, unfiltered adult fiction) and an explicit “not suited for” list (canon lore, instruction work, safety-critical, multilingual).

Usage — pull command, the exact /api/chat JSON with a filled persona card, and the one-line trace-stripping trick.

Training summary + ethics — corpus lineage, exclusions, optimizer, hardware, and adult-fiction-only scope.