The Athena-3 Family

Athena generated this model card! NOTE: This is only details for the 14B model. For other models, read their model page

Athena-3-14B is a 14.0-billion-parameter causal language model fine-tuned from Qwen2.5-14B-Instruct. This model is designed to provide highly fluent, contextually aware, and logically sound outputs across a broad range of NLP and reasoning tasks. It balances instruction-following with generative flexibility.

Model Details

Model Developer: Aayan Mishra
Model Type: Causal Language Model
Architecture: Transformer with Rotary Position Embeddings (RoPE), SwiGLU activation, RMSNorm, Attention QKV bias, and tied word embeddings
Parameters: 14.0 billion total (12.84 billion non-embedding)
Layers: 40
Attention Heads: 40 for query and 4 for key-value (Grouped Query Attention)
Vocabulary Size: Approximately 151,646 tokens
Context Length: Supports up to 131,072 tokens
Languages Supported: Over 29 languages, including strong performance in English, Chinese, and multilingual instruction tasks
License: MIT

Training Details

Athena-3-14B was fine-tuned using the Unsloth framework on a single NVIDIA T4 GPU. The fine-tuning process spanned approximately 90 minutes over 60 epochs, utilizing a curated instruction-tuned dataset. It is tailored for generalist NLP performance with a focus on reasoning, alignment, and fluency.

Intended Use

Athena-3-14B is ideal for a wide variety of tasks, including:

Instruction Following: Handling complex prompts with step-by-step logical output
Writing Assistance: Generating essays, emails, and coherent narratives
NLP Tasks: Summarization, question answering, translation, and text classification
STEM Support: Reasoning through academic and technical content

While Athena-3-14B is a versatile model, it is not intended for safety-critical applications or the handling of private, sensitive information.

How to Use

To utilize Athena-3-14B, ensure that you have the latest version of the transformers library installed:

pip install transformers

Here's an example of how to load the Athena-3-14B model and generate a response:

from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "Spestly/Athena-3-14B"
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
prompt = "Explain the concept of entropy in thermodynamics."
messages = [
    {"role": "system", "content": "You are Athena, an AI assistant designed to be helpful."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

Maverick Search usage 🔍

To use this model with Maverick Search, please refer to this repository

Limitations

Users should be aware of the following limitations:

Biases: Athena-3-14B may reflect biases from its pretraining and fine-tuning data. Outputs should be reviewed for fairness and accuracy.
Knowledge Cutoff: The model's knowledge is current as of August 2024.
Multilingual Performance: Performance varies by language, with strongest capabilities in English and aligned datasets.

Acknowledgements

Athena-3-14B builds upon the Qwen2.5-14B foundation. Special thanks to the open-source ecosystem and Unsloth for enabling efficient fine-tuning workflows.

License

Athena-3-14B is released under the MIT License, permitting broad use and distribution with proper attribution.

Contact

Email: athena@aayanmishra.com

Introducng Athena-3