Model Card for ambari-7b-lora-dora-v2

This model is a fine-tuned version of Cognitive-Lab/Ambari-7B-Instruct-v0.2. It has been trained using TRL.

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-lora-dora-v2", 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.

LoRA Configuration:

r (rank): 16 - Balances adaptation capacity with parameter efficiency lora_alpha: 16 - Scaling factor for LoRA updates lora_dropout: 0.0 - No dropout applied to LoRA layers target_modules: Applied to attention and feedforward projections: Query projections (q_proj) Key projections (k_proj) Value projections (v_proj) Output projections (o_proj) Gate projections (gate_proj) Up/Down projections (up_proj, down_proj) DoRA (Dimension-wise Ranking Adaptation) DoRA is enabled (use_dora: true) to further optimize the LoRA adaptation:

Decomposes weight updates into magnitude and direction components

Applies rank-restricted updates with improved generalization Reduces overfitting during CoT-specific fine-tuning Maintains baseline model's general knowledge while adapting for reasoning

Training Datasets

The model has been fine-tuned on diverse task-specific datasets from the base DoRA model:

Kanglish to English Translation: Translation capability from Kanglish to English English to Kanglish Translation: Translation capability from English to Kanglish

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}}
}