The smartest L3 8B model combined with high-end RP model. What could go wrong.
The idea was to fuse a bit of SimPO's realism with Stheno. It took a few days to come up with a balanced slerp configuration, but I'm more than satisfied with the end result.
SillyTavern
Text Completion presets
temp 0.9
top_k 30
top_p 0.75
min_p 0.2
rep_pen 1.1
smooth_factor 0.25
smooth_curve 1
Advanced Formatting
Context & Instruct preset by Virt-io
Instruct Mode: Enabled
merge
This is a merge of pre-trained language models created using mergekit.
This model was merged using the slerp merge method.
Models Merged
The following models were included in the merge:
Configuration
The following YAML configuration was used to produce this model:
slices:
- sources:
- model: Sao10K/L3-8B-Stheno-v3.2
layer_range: [0, 32]
- model: princeton-nlp/Llama-3-Instruct-8B-SimPO
layer_range: [0, 32]
merge_method: slerp
base_model: Sao10K/L3-8B-Stheno-v3.2
parameters:
t:
- filter: self_attn
value: [0.4, 0.5, 0.6, 0.4, 0.6]
- filter: mlp
value: [0.6, 0.5, 0.4, 0.6, 0.4]
- value: 0.5
dtype: bfloat16
Original model information:
Model: Sao10K/L3-8B-Stheno-v3.2
Stheno-v3.2-Zeta
Changes compared to v3.1 - Included a mix of SFW and NSFW Storywriting Data, thanks to Gryphe - Included More Instruct / Assistant-Style Data - Further cleaned up Roleplaying Samples from c2 Logs -> A few terrible, really bad samples escaped heavy filtering. Manual pass fixed it. - Hyperparameter tinkering for training, resulting in lower loss levels.
Testing Notes - Compared to v3.1 - Handles SFW / NSFW seperately better. Not as overly excessive with NSFW now. Kinda balanced. - Better at Storywriting / Narration. - Better at Assistant-type Tasks. - Better Multi-Turn Coherency -> Reduced Issues? - Slightly less creative? A worthy tradeoff. Still creative. - Better prompt / instruction adherence.
Want to support my work ? My Ko-fi page: https://ko-fi.com/tannedbum