You can deploy this model and start earning money today!

DazzlingShadow/
aya-expanse-32b-heretic-MPOA

Name: aya-expanse-32b-heretic-MPOA
Rating: 2.5 (10 reviews)

Views

No views yet

This is a decensored version of CohereLabs/aya-expanse-32b, made using Heretic v1.2.0 with Magnitude-Preserving Orthogonal Ablation (MPOA) enabled.

Abliteration parameters

Parameter	Value
direction_index	per layer
attn.o_proj.max_weight	1.28
attn.o_proj.max_weight_position	23.95
attn.o_proj.min_weight	0.77
attn.o_proj.min_weight_distance	17.10
mlp.down_proj.max_weight	1.42
mlp.down_proj.max_weight_position	27.61
mlp.down_proj.min_weight	1.35
mlp.down_proj.min_weight_distance	19.60

Performance

Metric	This model	Original model (CohereLabs/aya-expanse-32b)
KL divergence	0.2381	0 (by definition)
Refusals	13/100	99/100

Model Card for Aya-Expanse-32B

Aya Expanse 32B is an open-weight research release of a model with highly advanced multilingual capabilities. It focuses on pairing a highly performant pre-trained Command family of models with the result of a year’s dedicated research from Cohere Labs, including data arbitrage, multilingual preference training, safety tuning, and model merging. The result is a powerful multilingual large language model serving 23 languages.

This model card corresponds to the 32-billion version of the Aya Expanse model. We also released an 8-billion version which you can find here.

Developed by: Cohere Labs
Point of Contact: Cohere Labs
License: CC-BY-NC, requires also adhering to Cohere Lab's Acceptable Use Policy
Model: Aya Expanse 32B
Model Size: 32 billion parameters

Supported Languages

We cover 23 languages: Arabic, Chinese (simplified & traditional), Czech, Dutch, English, French, German, Greek, Hebrew, Hebrew, Hindi, Indonesian, Italian, Japanese, Korean, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Turkish, Ukrainian, and Vietnamese.

Try it: Aya Expanse in Action

Use the Cohere playground or our Hugging Face Space for interactive exploration.

How to Use Aya Expanse

Install the transformers library and load Aya Expanse 32B as follows:

python

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "CohereLabs/aya-expanse-32b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

# Format message with the chat template
messages = [{"role": "user", "content": "Anneme onu ne kadar sevdiğimi anlatan bir mektup yaz"}]
input_ids = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt")
## <BOS_TOKEN><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Anneme onu ne kadar sevdiğimi anlatan bir mektup yaz<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>

gen_tokens = model.generate(
    input_ids, 
    max_new_tokens=100, 
    do_sample=True, 
    temperature=0.3,
    )

gen_text = tokenizer.decode(gen_tokens[0])
print(gen_text)

Example Notebooks

Fine-Tuning:

Detailed Fine-Tuning Notebook.

Community-Contributed Use Cases::

The following notebooks contributed by Cohere Labs Community members show how Aya Expanse can be used for different use cases:

Model Details

Input: Models input text only.

Output: Models generate text only.

Model Architecture: Aya Expanse 32B is an auto-regressive language model that uses an optimized transformer architecture. Post-training includes supervised finetuning, preference training, and model merging.

Languages covered: The model is particularly optimized for multilinguality and supports the following languages: Arabic, Chinese (simplified & traditional), Czech, Dutch, English, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Korean, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Turkish, Ukrainian, and Vietnamese

Context length: 128K

Evaluation

We evaluated Aya Expanse 32B against Gemma 2 27B, Llama 3.1 70B, Mixtral 8x22B, and Qwen 2.5 32B using the dolly_human_edited subset from the Aya Evaluation Suite dataset and m-ArenaHard, a dataset based on the Arena-Hard-Auto dataset and translated to the 23 languages we support in Aya Expanse. Win-rates were determined using gpt-4o-2024-08-06 as a judge. For a conservative benchmark, we report results from gpt-4o-2024-08-06, though gpt-4o-mini scores showed even stronger performance.

The m-ArenaHard dataset, used to evaluate Aya Expanse’s capabilities, is publicly available here.

Model Card Contact

For errors or additional questions about details in this model card, contact labs@cohere.com

Terms of Use

We hope that the release of this model will make community-based research efforts more accessible, by releasing the weights of a highly performant multilingual model to researchers all over the world. This model is governed by a CC-BY-NC, requires also adhering to Cohere Lab's Acceptable Use Policy

Cite

You can cite Aya Expanse using:

@misc{dang2024ayaexpansecombiningresearch,
      title={Aya Expanse: Combining Research Breakthroughs for a New Multilingual Frontier}, 
      author={John Dang and Shivalika Singh and Daniel D'souza and Arash Ahmadian and Alejandro Salamanca and Madeline Smith and Aidan Peppin and Sungjin Hong and Manoj Govindassamy and Terrence Zhao and Sandra Kublik and Meor Amer and Viraat Aryabumi and Jon Ander Campos and Yi-Chern Tan and Tom Kocmi and Florian Strub and Nathan Grinsztajn and Yannis Flet-Berliac and Acyr Locatelli and Hangyu Lin and Dwarak Talupuru and Bharat Venkitesh and David Cairuz and Bowen Yang and Tim Chung and Wei-Yin Ko and Sylvie Shang Shi and Amir Shukayev and Sammie Bae and Aleksandra Piktus and Roman Castagné and Felipe Cruz-Salinas and Eddie Kim and Lucas Crawhall-Stein and Adrien Morisot and Sudip Roy and Phil Blunsom and Ivan Zhang and Aidan Gomez and Nick Frosst and Marzieh Fadaee and Beyza Ermis and Ahmet Üstün and Sara Hooker},
      year={2024},
      eprint={2412.04261},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.04261}, 
}

DazzlingShadow/aya-expanse-32b-heretic-MPOA