docs: honest attribution — base_model Qwen/Qwen3-8B + license + credit; drop fabricated 32B/Zen-3-arch/2024 claims
Browse files
README.md
CHANGED
|
@@ -11,23 +11,25 @@ tags:
|
|
| 11 |
- CJK
|
| 12 |
pipeline_tag: text-generation
|
| 13 |
library_name: transformers
|
| 14 |
-
base_model:
|
| 15 |
---
|
| 16 |
|
| 17 |
# Zen Multilingual
|
| 18 |
|
| 19 |
-
|
|
|
|
|
|
|
| 20 |
|
| 21 |
Multilingual generation across 30+ languages: English, Chinese, Japanese, Korean, Arabic, Spanish, French, German, Portuguese, Russian, and more.
|
| 22 |
|
| 23 |
Strong at cross-lingual reasoning, code-switching, and multilingual instruction following.
|
| 24 |
|
| 25 |
-
Base weights: [
|
| 26 |
|
| 27 |
```python
|
| 28 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 29 |
-
model = AutoModelForCausalLM.from_pretrained("zenlm/zen-
|
| 30 |
-
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-
|
| 31 |
messages = [{"role": "user", "content": "Your domain-specific prompt here"}]
|
| 32 |
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 33 |
inputs = tokenizer(text, return_tensors="pt").to(model.device)
|
|
@@ -35,6 +37,10 @@ output = model.generate(**inputs, max_new_tokens=1024)
|
|
| 35 |
print(tokenizer.decode(output[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))
|
| 36 |
```
|
| 37 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
---
|
| 39 |
## The Zen LM Family
|
| 40 |
|
|
|
|
| 11 |
- CJK
|
| 12 |
pipeline_tag: text-generation
|
| 13 |
library_name: transformers
|
| 14 |
+
base_model: Qwen/Qwen3-8B
|
| 15 |
---
|
| 16 |
|
| 17 |
# Zen Multilingual
|
| 18 |
|
| 19 |
+
Fine-tuned from [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) (Apache 2.0) with Hanzo identity + agentic-data training + abliteration.
|
| 20 |
+
|
| 21 |
+
> **Base**: Qwen3-8B | **Parameters**: 8B | **Architecture**: Qwen3 | **Context**: 128K | **License**: Apache 2.0
|
| 22 |
|
| 23 |
Multilingual generation across 30+ languages: English, Chinese, Japanese, Korean, Arabic, Spanish, French, German, Portuguese, Russian, and more.
|
| 24 |
|
| 25 |
Strong at cross-lingual reasoning, code-switching, and multilingual instruction following.
|
| 26 |
|
| 27 |
+
Base weights: [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) (Apache 2.0).
|
| 28 |
|
| 29 |
```python
|
| 30 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 31 |
+
model = AutoModelForCausalLM.from_pretrained("zenlm/zen-multilingual", torch_dtype="auto")
|
| 32 |
+
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-multilingual")
|
| 33 |
messages = [{"role": "user", "content": "Your domain-specific prompt here"}]
|
| 34 |
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 35 |
inputs = tokenizer(text, return_tensors="pt").to(model.device)
|
|
|
|
| 37 |
print(tokenizer.decode(output[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))
|
| 38 |
```
|
| 39 |
|
| 40 |
+
## Credits
|
| 41 |
+
|
| 42 |
+
Built on [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) by the Qwen Team, Alibaba Cloud, licensed under Apache 2.0. Hanzo adds identity training, agentic-data fine-tuning, and abliteration.
|
| 43 |
+
|
| 44 |
---
|
| 45 |
## The Zen LM Family
|
| 46 |
|