[!NOTE] The LFM3 1.7B version, Schauspieler 2B, may be published tomorrow if I like it. The first training run had a good enough eval score but it didn't feel good enough, so I'm trying different hyperparameters.
Support me here if you're interested:
ko-fi.com/hamzah0asadullah
Ein Schauspieler ist jemand, der u.a. mithilfe von Sprache eine fiktive Rolle zur Schau spielt.
Image by kinoko no sato from Danbooru.
Schauspieler 1B is a finetune on Qwen/Qwen3.5-0.8B using Gryphe/Sonnet3.5-Charcard-Roleplay. Trained using a humble AMD RX 7900 XT for ~26h over three epochs (mean token accuracy on eval from ~0.5 to ~0.7).
whisper
You'll notice the small size quickly. LFM3-1.7B version WIP
For wishes, just open a discussion here or ping me on (KoboldAI) Discord @hamzah0asadullah.
Generation parameters copied from upstream, use temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0.
Thinking may or may not work, see what works and what doesn't. Chat format also from upstream, use ChatML:
<|im_start|>system
You are abc in this endless, qualitative roleplay.
abc is a 24-year old working-class man who lives alone in a mansion in Bavaria, Germany.
{how abc behaves and interacts with {{user}}.}
{how abc looks visually}
{abc lore}<|im_end|>
<|im_start|>assistant
{first utterance/scenario}<|im_end|>
<|im_start|>user
*actions in asterisks* "Things you say in quotes"<|im_end|>