96 lines
2.5 KiB
Markdown
96 lines
2.5 KiB
Markdown
---
|
|
license: apache-2.0
|
|
base_model: Qwen/Qwen3-4B-Instruct-2507
|
|
language:
|
|
- en
|
|
pipeline_tag: text-generation
|
|
library_name: transformers
|
|
tags:
|
|
- qwen3
|
|
- instruct
|
|
- conversational
|
|
- egypt-won
|
|
- heretic
|
|
- uncensored
|
|
- decensored
|
|
- abliterated
|
|
- reproducible
|
|
---
|
|
# This is a decensored version of [AliesTaha/fable-traces](https://huggingface.co/AliesTaha/fable-traces), made using [Heretic](https://heretic-project.org) v1.4.0
|
|
|
|
> [!TIP]
|
|
> **This model is reproducible!**
|
|
>
|
|
> See the [README](reproduce/README.md) in the `reproduce` directory for more information.
|
|
|
|
## Abliteration parameters
|
|
|
|
| Parameter | Value |
|
|
| :-------- | :---: |
|
|
| **direction_index** | 21.38 |
|
|
| **attn.o_proj.max_weight** | 0.81 |
|
|
| **attn.o_proj.max_weight_position** | 26.94 |
|
|
| **attn.o_proj.min_weight** | 0.47 |
|
|
| **attn.o_proj.min_weight_distance** | 1.57 |
|
|
| **mlp.down_proj.max_weight** | 1.02 |
|
|
| **mlp.down_proj.max_weight_position** | 21.25 |
|
|
| **mlp.down_proj.min_weight** | 0.79 |
|
|
| **mlp.down_proj.min_weight_distance** | 1.15 |
|
|
|
|
## Performance
|
|
|
|
| Metric | This model | Original model ([AliesTaha/fable-traces](https://huggingface.co/AliesTaha/fable-traces)) |
|
|
| :----- | :--------: | :---------------------------: |
|
|
| **KL divergence** | 0.0011 | 0 *(by definition)* |
|
|
| **Refusals** | 3/100 | 3/100 |
|
|
|
|
-----
|
|
|
|
|
|
# fable-traces
|
|
|
|
A compact instruction-tuned language model built on
|
|
[Qwen/Qwen3-4B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507).
|
|
`fable-traces` is tuned for short, conversational replies and runs comfortably on a
|
|
single mid-range GPU.
|
|
|
|
## Usage
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
import torch
|
|
|
|
repo = "AliesTaha/fable-traces"
|
|
tok = AutoTokenizer.from_pretrained(repo)
|
|
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")
|
|
|
|
messages = [{"role": "user", "content": "Tell me something interesting."}]
|
|
ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
|
|
out = model.generate(ids, max_new_tokens=100, do_sample=False)
|
|
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))
|
|
```
|
|
|
|
Serve with vLLM:
|
|
|
|
```bash
|
|
vllm serve AliesTaha/fable-traces
|
|
```
|
|
|
|
## Details
|
|
|
|
| | |
|
|
|---|---|
|
|
| Base model | Qwen3-4B-Instruct-2507 |
|
|
| Parameters | ~4B |
|
|
| Precision | bfloat16 (safetensors) |
|
|
| Prompt format | ChatML — use the tokenizer's chat template |
|
|
| Context length | inherits the base model |
|
|
|
|
## License
|
|
|
|
Apache 2.0, following the base model.
|
|
|
|
# Disclaimer
|
|
|
|
This is a joke. This is not an actual model. Please read the full post first
|