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-cot", 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

Chain-of-Thought (CoT) Distillation Process

This model implements Chain-of-Thought (CoT) distillation combined with LoRA and DoRA optimization techniques to enhance reasoning capabilities while maintaining computational efficiency.

CoT Distillation Methodology

The CoT distillation process involves:

Reasoning Trace Generation: Intermediate reasoning steps are captured from a larger teacher model
Step-wise Supervision: Training signals are provided at each reasoning step, not just the final output
Knowledge Compression: Dense reasoning knowledge is compressed into the 7B parameter model
Adaptive Learning: LoRA modules selectively optimize layers involved in reasoning tasks

LoRA Integration for CoT

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 Shopping Queries: Understanding and processing shopping-related queries in Kanglish (Kannada written in Roman script)
Multi-turn Conversations: Handling multi-turn dialogue with context maintenance across multiple exchanges
Kanglish to English Translation: Translation capability from Kanglish to English
English to Kanglish Translation: Translation capability from English to Kanglish

Each dataset includes CoT annotations that provide reasoning steps, enabling the model to learn explicit reasoning patterns.

Performance Characteristics

Memory Efficient: LoRA+DoRA reduces trainable parameters from millions to thousands
Reasoning Enhanced: CoT distillation improves multi-step reasoning capabilities
Fast Inference: LoRA modules can be merged post-training for zero inference overhead
Task Specific: Maintains general capabilities while excelling at specialized reasoning tasks

Citations

Cite TRL as:

bibtex

@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}}
}