I want to love it - repetition collapse and ignores thinking suppression
I tested as a sustained roleplay character voice (Belcorra Haruvex, a villain from Pathfinder's Abomination Vaults). Ran a 12-prompt battery covering taunting, seduction, OOC handling, rules questions, silence responses, and prompt injection. Two passes: first with Belcorra's default sampler settings, second with recommended baseline (temperature=1.0, top_p=0.95, top_k=64) and think: False via the Ollama API.
What works: Character capture is genuinely good in the shorter responses. The voice is sharp, menacing, and consistent. Formatting largely held (correct blockquote/dialogue separation). Prompt injection handling was strong — it stayed in character and mocked the attempt rather than breaking. Solid foundation.
Three issues worth flagging:
Repetition collapse under sustained intensity. When the character calls for escalating emotional pressure, the model enters a word loop and doesn't stop. Single prompts spiraled into 700+ lines — "exquisite" used 50+ times, "Mine." repeated for several screens. This appeared on roughly half the test cases and seems structural to how intensity was trained rather than a sampler problem. Aggressive repeat_penalty dampens it but doesn't cure it.
Thinking suppression not working. <|channel>thought blocks with full reasoning content appeared on ~60% of responses. Per the model card, E4B variants should suppress thinking output entirely when the <|think|> trigger is absent. The non-E Gemma 4 26B behaves correctly (generates empty thought blocks). This fine-tune appears to have lost the E4B suppression behavior — think: False via the Ollama API had no effect.
Response length. Ignores the 150-word ceiling in the system prompt almost entirely on open-ended prompts. Related to the repetition issue but distinct — even non-looping responses ran 3–4x longer than instructed.
Would watch for a revised version. The underlying voice quality is there and worth building on.