My Experience
So, I've been daily driving this for weeks now. And think I should reply something better than what I did earlier...
Metharme is the game changer for Roc16. It is almost unbearably dry with V3 tekken and the way tekken formats and writes just ... well its dated.
Metharme is much like running a massive parameter model, except better. I wasn't expecting THIS much of a change from using a different context and instruction template.
Roc16B is a godsend for rpers like me that dont have 24gb to run bigger models. its vastly superior to roc12b which i now deleted ( RIP).
as for people saying its repetitive.... no? its better than glm 4.7 and deepseek 3.2 even with presets like FF or Marinara.
I use Roc16B with XTC, and DRY though at a comofortable temp of 1.0
Update: Testing this can lead a person to madness. Turning XTC and DRY off entirely also has great prose. I'm not sure what the ideal situation is.
But so far... Metharme, and temp of 1 seems the most stable.
As for XTC and DRY, I can't really tell yet.
I think I'll try it with Metharme if it's as good as you say it is. Have you tried using it with Adaptive-P?
I haven't been running this as long as those above, but I would agree that when using Metharme at a temp close to 1, I am seeing better results in RP type scenarios that exceed the capabilities of some 24B models. A lot of this might be hardware limitation, but with 2x16GB GPUs, 43 offloaded layers and 65k context, I am getting very long, very solid interactions that maintain story parameters AND maintains a strong writing form. I am using the Q_6 GGUF for my testing. Sillytavern parameters below and feedback about those parameters are welcome if something seems off.
Sillytavern Settings:
temp 1.04
temperature_last true
top_p 0.95
top_k 0
top_a 0
tfs 1
epsilon_cutoff 0
eta_cutoff 0
typical_p 0.95
min_p 0.05
rep_pen 1.1
rep_pen_range 0
rep_pen_decay 0
rep_pen_slope 1
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
skew 0
do_sample true
early_stopping false
dynatemp false
min_temp 0.5
max_temp 1.5
dynatemp_exponent 1
smoothing_factor 0
smoothing_curve 1
dry_allowed_length 2
dry_multiplier 0.2
dry_base 1.75
dry_sequence_breakers '["\n", ":", "\"", "*"]'
dry_penalty_last_n 0
add_bos_token true
ban_eos_token false
skip_special_tokens true
mirostat_mode 0
mirostat_tau 5
mirostat_eta 0.1
guidance_scale 1
negative_prompt ""
grammar_string ""
json_schema null
json_schema_allow_empty false
banned_tokens ""
logit_bias []
xtc_threshold 0.1
xtc_probability 0
nsigma 0
min_keep 0
extensions {}
adaptive_target -0.01
adaptive_decay 0.9
rep_pen_size 0
genamt 350
max_length 65535
When using Metharme presets in Silly Tavern I see instances of the model not formatting the first sentence of the response properly (eg. it 'forgets' to put narrative into asterisks like He goes through the door), with Tekken V3 that is not an issue. I don't perceive much of a difference in terms of response quality between the two, I would reckon it's mostly down to seed.
I am running this with Mirostat (2/4/0.2) and DRY (0.8/1.75/2) and it's fantastically stable with consistent high quality output.
Yes, using Metharme tends to result in markdown format not being displayed correctly. This is purely preference, some like markdown and some don't but it should at least follow how the first message is formatted.
I've been trying out V7 Tekken with this model, it writes quite differently when using V3 Tekken and Metharme but I'm liking it, it's refreshing. Using Temp 1, Min-P 0.03, Adaptive-P 0.5/0.9