|
|
--- |
|
|
license: apache-2.0 |
|
|
base_model: |
|
|
- Qwen/Qwen3-14B |
|
|
- huihui-ai/Qwen3-14B-abliterated |
|
|
- Qwen/Qwen3-14B-Base |
|
|
library_name: transformers |
|
|
tags: |
|
|
- mergekit |
|
|
- merge |
|
|
- qwen3 |
|
|
- uncensored |
|
|
- reasoning |
|
|
--- |
|
|
# Qwen3-14B-abliterated-TIES |
|
|
|
|
|
This is a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit). |
|
|
|
|
|
## Merge Details |
|
|
### Merge Method |
|
|
|
|
|
This model was merged using the [TIES](https://arxiv.org/abs/2306.01708) merge method using [Qwen/Qwen3-14B-Base](https://huggingface.co/Qwen/Qwen3-14B-Base) as a base. |
|
|
|
|
|
### Models Merged |
|
|
|
|
|
The following models were included in the merge: |
|
|
* [huihui-ai/Qwen3-14B-abliterated](https://huggingface.co/huihui-ai/Qwen3-14B-abliterated) |
|
|
|
|
|
### Configuration |
|
|
|
|
|
The following YAML configuration was used to produce this model: |
|
|
|
|
|
```yaml |
|
|
models: |
|
|
- model: huihui-ai/Qwen3-14B-abliterated |
|
|
parameters: |
|
|
weight: 1 |
|
|
density: 1 |
|
|
merge_method: ties |
|
|
base_model: Qwen/Qwen3-14B-Base |
|
|
parameters: |
|
|
weight: 1 |
|
|
density: 1 |
|
|
normalize: true |
|
|
int8_mask: true |
|
|
dtype: bfloat16 |
|
|
|
|
|
|
|
|
``` |
|
|
|
|
|
## Reasoning Fix |
|
|
|
|
|
The abliteration and merge caused an issue where the `<think>` token would not always be properly selected. This was fixed by using the vector from [Qwen/Qwen3-14B](https://huggingface.co/Qwen/Qwen3-14B). |
|
|
|
|
|
```python |
|
|
import torch |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
|
|
|
# paths |
|
|
src = "Qwen/Qwen3-14B" |
|
|
tgt = "TARGET_MODEL" |
|
|
out = "OUTPUT_DIR" |
|
|
|
|
|
tok_tag = "<think>" |
|
|
|
|
|
# load |
|
|
src_tok = AutoTokenizer.from_pretrained(src) |
|
|
tgt_tok = AutoTokenizer.from_pretrained(tgt) |
|
|
src_model = AutoModelForCausalLM.from_pretrained(src, torch_dtype="auto", device_map="cpu") |
|
|
tgt_model = AutoModelForCausalLM.from_pretrained(tgt, torch_dtype="auto", device_map="cpu") |
|
|
|
|
|
# ids (don’t hard-code, trust the tokenizer) |
|
|
sid = src_tok.convert_tokens_to_ids(tok_tag) |
|
|
tid = tgt_tok.convert_tokens_to_ids(tok_tag) |
|
|
|
|
|
if tid == src_tok.unk_token_id: |
|
|
# tgt lost the token – add it back, resize, grab new id |
|
|
tgt_tok.add_tokens([tok_tag]) |
|
|
tid = tgt_tok.convert_tokens_to_ids(tok_tag) |
|
|
tgt_model.resize_token_embeddings(len(tgt_tok)) |
|
|
|
|
|
# copy the vec |
|
|
with torch.no_grad(): |
|
|
tgt_model.get_input_embeddings().weight[tid].copy_( |
|
|
src_model.get_input_embeddings().weight[sid] |
|
|
) |
|
|
|
|
|
# optional blend instead of overwrite |
|
|
# tgt_vec = tgt_model.get_input_embeddings().weight[tid] |
|
|
# tgt_model.get_input_embeddings().weight[tid].copy_(0.7*src_vec + 0.3*tgt_vec) |
|
|
|
|
|
# save |
|
|
tgt_model.save_pretrained(out) |
|
|
tgt_tok.save_pretrained(out) |
|
|
|
|
|
``` |
