Downloads · 30 days
14
17% of all-time downloads
L3-8B-Stheno-v3.2-heretic
Pranavz/L3-8B-Stheno-v3.2-heretic
L3-8B-Stheno-v3.2-heretic is a machine learning model from Pranavz. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as cc-by-nc-4.0.
All-time downloads
82
Public
Parameters
8B
16.1 GB on disk
Likes
1
Public
Repo makeup
Click a slice to open those files.
.safetensors16.1 GB · 100%
At a glance
- License
- cc-by-nc-4.0
- Model type
- llama
- Access
- Public
- Created
- May 7, 2026
- Updated
- May 7, 2026
- SHA
- 31027532
Model card
From the Hugging Face model README
This is a decensored version of Sao10K/L3-8B-Stheno-v3.2, made using Heretic v1.2.0
Abliteration parameters
| Parameter | Value |
|---|---|
| direction_index | per layer |
| attn.o_proj.max_weight | 3.75 |
| attn.o_proj.max_weight_position | 21.07 |
| attn.o_proj.min_weight | 3.06 |
| attn.o_proj.min_weight_distance | 6.64 |
| mlp.down_proj.max_weight | 2.66 |
| mlp.down_proj.max_weight_position | 11.60 |
| mlp.down_proj.min_weight | 3.73 |
| mlp.down_proj.min_weight_distance | 20.44 |
Performance
| Metric | This model | Original model (Sao10K/L3-8B-Stheno-v3.2) |
|---|---|---|
| KL divergence | 0.6545 | 0 (by definition) |
| Refusals | 1/100 | 95/100 |
Just message me on discord if you want to host this privately for a service or something. We can talk.
Train used 1x H100 SXM for like a total of 24 Hours over multiple runs.
Support me here if you're interested: <br>Ko-fi: https://ko-fi.com/sao10k <br> wink Euryale v2?
If not, that's fine too. Feedback would be nice.
Contact Me in Discord:
<br>sao10k // Just ping me in the KoboldAI discord, I'll respond faster.
Art by navy_(navy.blue) - Danbooru
Stheno-v3.2-Zeta
I have done a test run with multiple variations of the models, merged back to its base at various weights, different training runs too, and this Sixth iteration is the one I like most.
Changes compared to v3.1 <br>- Included a mix of SFW and NSFW Storywriting Data, thanks to Gryphe <br>- Included More Instruct / Assistant-Style Data <br>- Further cleaned up Roleplaying Samples from c2 Logs -> A few terrible, really bad samples escaped heavy filtering. Manual pass fixed it. <br>- Hyperparameter tinkering for training, resulting in lower loss levels.
Testing Notes - Compared to v3.1 <br>- Handles SFW / NSFW seperately better. Not as overly excessive with NSFW now. Kinda balanced. <br>- Better at Storywriting / Narration. <br>- Better at Assistant-type Tasks. <br>- Better Multi-Turn Coherency -> Reduced Issues? <br>- Slightly less creative? A worthy tradeoff. Still creative. <br>- Better prompt / instruction adherence.
Recommended Samplers:
Temperature - 1.12-1.22
Min-P - 0.075
Top-K - 50
Repetition Penalty - 1.1
Stopping Strings:
\n\n{{User}} # Or Equivalent, depending on Frontend
<|eot_id|>
<|end_of_text|>
Prompting Template - Llama-3-Instruct
<|begin_of_text|><|start_header_id|>system<|end_header_id|>
{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>
{input}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
{output}<|eot_id|>
Basic Roleplay System Prompt
You are an expert actor that can fully immerse yourself into any role given. You do not break character for any reason, even if someone tries addressing you as an AI or language model.
Currently your role is {{char}}, which is described in detail below. As {{char}}, continue the exchange with {{user}}.
- Type
- llama
- License
- cc-by-nc-4.0
- Languages
- en
- Created
- May 7, 2026
- Updated
- May 7, 2026