Model Card for ambari-7b-dora

This model is a fine-tuned version of Cognitive-Lab/Ambari-7B-Instruct-v0.2 fine-tuned using the DoRA (Dimension-wise Ranking Adaptation) technique.

Fine-tuning Details

This model has been fine-tuned on the following tasks and datasets:

• Kanglish Shopping Queries: Fine-tuned to understand and process shopping-related queries in Kanglish (Kannada written in Roman script)
• Multi-turn Conversations: Optimized for handling multi-turn dialogue and maintaining context across multiple exchanges
• Kanglish to English Translation: Trained to translate from Kanglish to English
• English to Kanglish Translation: Trained to translate from English to Kanglish

Fine-tuning Technique

The model uses the DoRA (Dimension-wise Ranking Adaptation) technique, which is an efficient fine-tuning approach that adapts specific dimensions of the model weights rather than fine-tuning all parameters. This approach:

• Reduces memory requirements during training
• Maintains the base model's capabilities while adding task-specific knowledge
• Provides efficient adaptation for multiple downstream tasks

Quick start

from transformers import pipeline

question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
generator = pipeline("text-generation", model="Akshaymp/ambari-7b-dora", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])

Training procedure

This model was trained with SFT.

Framework versions

• TRL: 0.24.0
• Transformers: 4.57.3
• Pytorch: 2.9.1
• Datasets: 4.3.0
• Tokenizers: 0.22.1

Citations

Cite TRL as:

@misc{vonwerra2022trl,
    title        = {{TRL: Transformer Reinforcement Learning}},
    author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
    year         = 2020,
    journal      = {GitHub repository},
    publisher    = {GitHub},
    howpublished = {\url{https://github.com/huggingface/trl}}
}