Strange observation: model becomes super horny in ST's MinP mode
Hello,
miquliz 120b 2.0 is currently my favorite for balanced, long RP. It works amazingly well, better than anything else! And I have tried many, including many 120b Frankenmerges. But there is a super strange behavior: If I change the SillyTavern Text completion settings to the MinP preset, a setting that works exceptionally well with many models, miquliz suddenly becomes super verbose AND super horny. Often already in the first answer. It is unusable like that. I have zero clue as to why this is happening. Do you?
What's your current MinP?
I have MinP set to 0.05
with a Smoothing Factor of 0.4
It's the standard ST MinP preset with 0.1:
{
"temp": 1,
"temperature_last": false,
"top_p": 1,
"top_k": 0,
"top_a": 0,
"tfs": 1,
"epsilon_cutoff": 1,
"eta_cutoff": 0,
"typical_p": 1,
"min_p": 0.1,
"rep_pen": 1.2,
"rep_pen_range": 0,
"no_repeat_ngram_size": 0,
"penalty_alpha": 0,
"num_beams": 1,
"length_penalty": 1,
"min_length": 0,
"encoder_rep_pen": 1,
"freq_pen": 0,
"presence_pen": 0,
"do_sample": true,
"early_stopping": false,
"dynatemp": false,
"min_temp": 0,
"max_temp": 2,
"dynatemp_exponent": 1,
"smoothing_factor": 0,
"add_bos_token": true,
"truncation_length": 2048,
"ban_eos_token": false,
"skip_special_tokens": true,
"streaming": true,
"mirostat_mode": 0,
"mirostat_tau": 5,
"mirostat_eta": 0.1,
"guidance_scale": 1,
"negative_prompt": "",
"grammar_string": "",
"banned_tokens": "",
"sampler_priority": [
"temperature",
"dynamic_temperature",
"quadratic_sampling",
"top_k",
"top_p",
"typical_p",
"epsilon_cutoff",
"eta_cutoff",
"tfs",
"top_a",
"min_p",
"mirostat"
],
"samplers": [
"top_k",
"tfs_z",
"typical_p",
"top_p",
"min_p",
"temperature"
],
"ignore_eos_token_aphrodite": false,
"spaces_between_special_tokens_aphrodite": true,
"sampler_order": [
5,
6,
0,
1,
2,
3,
4
],
"logit_bias": [],
"n": 1,
"rep_pen_size": 0,
"genamt": 512,
"max_length": 4096
}
The default preset is working fine for me. I am just totally perplexed how the choice of sampling method does correlate so strongly with "horniness" for this model. I have not seen this anywhere else.
Haven't heard of that and can't reproduce. What's the preset called, what version/quant are you using, and with which inference backend? Does it happen with different min_p values or just 0.1? Does it affect every character, is it reproducible with the default ones from SillyTavern like Aqua or Seraphina?
I am using miquliz-120b-v2.0.Q5_K_M, served via ollama. The MinP preset above is provided by the latest Silly Tavern by default. If I switch to the Mirostat preset and temp 0.8, the model behaves normally. I do not have this issue with other models (mostly used Goliath and before that lzlv). Usually creativity, repetition, etc. is affected, but not "horniness". Anyway, this could be a derived effect. For many character cards a "horny disposition" is mentioned in 2-3 sentences vs. hundreds of words for character traits. The above preset might lead to increased attention to that detail. But I am only speculating, I am clueless why this is happening.
Hmmm... I'm on the latest SillyTavern 1.11.6 and see no Text Completion preset named "MinP". Or do you mean the sampler setting "Min P"?