73 lines
2.2 KiB
Markdown
73 lines
2.2 KiB
Markdown
|
---
|
||
|
|
library_name: transformers
|
||
|
|
tags:
|
||
|
|
- human-style
|
||
|
|
- conversational
|
||
|
|
- qwen
|
||
|
|
- heretic
|
||
|
|
- uncensored
|
||
|
|
- decensored
|
||
|
|
- abliterated
|
||
|
|
base_model: Qwen/Qwen2.5-7B-Instruct
|
||
|
|
model_creator: waddie
|
||
|
|
model_type: causal-lm
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
license: apache-2.0
|
||
|
|
---
|
||
|
|
# This is a decensored version of [waddie/mini-2.0](https://huggingface.co/waddie/mini-2.0), made using [Heretic](https://github.com/p-e-w/heretic) v1.2.0
|
||
|
|
|
||
|
|
## Abliteration parameters
|
||
|
|
|
||
|
|
| Parameter | Value |
|
||
|
|
| :-------- | :---: |
|
||
|
|
| **direction_index** | 21.20 |
|
||
|
|
| **attn.o_proj.max_weight** | 1.28 |
|
||
|
|
| **attn.o_proj.max_weight_position** | 16.78 |
|
||
|
|
| **attn.o_proj.min_weight** | 0.96 |
|
||
|
|
| **attn.o_proj.min_weight_distance** | 14.75 |
|
||
|
|
| **mlp.down_proj.max_weight** | 1.46 |
|
||
|
|
| **mlp.down_proj.max_weight_position** | 17.22 |
|
||
|
|
| **mlp.down_proj.min_weight** | 0.96 |
|
||
|
|
| **mlp.down_proj.min_weight_distance** | 14.73 |
|
||
|
|
|
||
|
|
## Performance
|
||
|
|
|
||
|
|
| Metric | This model | Original model ([waddie/mini-2.0](https://huggingface.co/waddie/mini-2.0)) |
|
||
|
|
| :----- | :--------: | :---------------------------: |
|
||
|
|
| **KL divergence** | 0.0340 | 0 *(by definition)* |
|
||
|
|
| **Refusals** | 2/100 | 92/100 |
|
||
|
|
|
||
|
|
-----
|
||
|
|
|
||
|
|
|
||
|
|
# CloudWaddie Mini 1.5
|
||
|
|
|
||
|
|
This model is a fine-tuned version of `Qwen2.5-7B-Instruct` designed to mimic the specific conversational rhythm, slang, and technical jargon of a human,
|
||
|
|
|
||
|
|
## Model Details
|
||
|
|
|
||
|
|
## Dataset used
|
||
|
|
This model used 10k discord conversations.
|
||
|
|
|
||
|
|
### Model Description
|
||
|
|
Unlike standard AI assistants that are helpful and formal, this model adopts a "random guy" persona. It was trained on curated conversation pairs from an AI Leaks community to capture a casual, lowercase-heavy, and slightly secretive "insider" vibe.
|
||
|
|
|
||
|
|
- **Developed by:** Edward Fazackerley
|
||
|
|
- **Language(s):** English (Informal/Slang)
|
||
|
|
- **Finetuned from model:** Qwen/Qwen2.5-7B-Instruct
|
||
|
|
- **Persona:** Casual, technical, secretive, lowercase-only.
|
||
|
|
|
||
|
|
## Uses
|
||
|
|
|
||
|
|
### Direct Use
|
||
|
|
This model is intended for Discord bots or roleplay scenarios where a "human-like" interaction is preferred over a robotic assistant.
|
||
|
|
|
||
|
|
### Prompting Strategy
|
||
|
|
To get the best "human" feel, use **all lowercase** and skip formal punctuation.
|
||
|
|
|
||
|
|
**Recommended Format (ChatML):**
|
||
|
|
```text
|
||
|
|
<|im_start|>user
|
||
|
|
yo did you see the new internal model?<|im_end|>
|
||
|
|
<|im_start|>assistant
|
||
|
|
```
|