vedai-001-3B

A fine-tuned version of Llama-3.2-3B-Instruct trained on reflective reasoning tasks, designed to mimic human stream-of-consciousness thinking with thorough, iterative reasoning.

Model Details

Model Description

This model is a LoRA fine-tuned adapter based on Llama-3.2-3B-Instruct, specifically trained to engage in reflective, step-by-step reasoning. It emphasizes exploration, self-doubt, and continuous refinement before arriving at conclusions, similar to human thought processes.

• Developed by: Harsh Bopaliya
• Model type: Causal Language Model (LoRA Adapter)
• Language(s): English
• License: Llama 3.2 Community License
• Finetuned from model: unsloth/Llama-3.2-3B-Instruct
• Training Framework: Unsloth + PEFT

Model Sources

• Base Model: unsloth/Llama-3.2-3B-Instruct
• Training Dataset: ServiceNow-AI/R1-Distill-SFT (v0)

Uses

Direct Use

This model is designed for tasks requiring deep reasoning and thoughtful problem-solving:

• Complex problem solving with step-by-step reasoning
• Mathematical and logical reasoning tasks
• Reflective thinking and analysis
• Educational tutoring with detailed explanations

Downstream Use

Can be integrated into applications requiring:

• AI assistants with enhanced reasoning capabilities
• Educational platforms
• Research tools requiring chain-of-thought reasoning

Out-of-Scope Use

• Real-time decision making in critical systems
• Medical diagnosis or legal advice
• Tasks requiring factual accuracy without reasoning verification
• Production systems without human oversight

Bias, Risks, and Limitations

• Inherits biases from the base Llama-3.2-3B model and training data
• May generate overly verbose responses due to reflective reasoning training
• Reasoning quality depends on problem complexity and domain
• Limited to the knowledge cutoff of the base model

Recommendations

Users should:

• Verify critical information from reliable sources
• Be aware that reasoning steps may not always be logically sound
• Use appropriate safeguards for production deployments
• Monitor outputs for bias and factual accuracy

How to Get Started with the Model

from unsloth import FastLanguageModel
import torch

# Load model and tokenizer
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="your-username/vedai-001-3B",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)

# Enable native 2x faster inference
FastLanguageModel.for_inference(model)

# Prepare prompt
prompt = """You are a reflective assistant engaging in thorough, iterative reasoning, mimicking human stream-of-consciousness thinking.
<problem>
What is 15% of 240?
</problem>"""

inputs = tokenizer([prompt], return_tensors="pt").to("cuda")

# Generate
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    use_cache=True,
    temperature=0.7,
    top_p=0.9,
)

print(tokenizer.batch_decode(outputs))

Training Details

Training Data

The model was fine-tuned on the ServiceNow-AI/R1-Distill-SFT dataset (v0), which contains:

• Problem statements
• Reannotated assistant reasoning (stream-of-consciousness style)
• Final solutions

The data was formatted using a custom prompt template emphasizing reflective, iterative reasoning.

Training Procedure

Preprocessing

Data was formatted using the following prompt structure:

r1_prompt = """You are a reflective assistant engaging in thorough, iterative reasoning, mimicking human stream-of-consciousness thinking. Your approach emphasizes exploration, self-doubt, and continuous refinement before coming up with an answer.
<problem>
{}
</problem>
{}
{}
"""

Training Hyperparameters

• Training regime: Mixed precision (fp16/bf16)
• Optimizer: AdamW 8-bit
• Learning rate: 2e-4
• Batch size per device: 2
• Gradient accumulation steps: 4
• Max steps: 60
• Warmup steps: 5
• Weight decay: 0.01
• LR scheduler: Linear
• Max sequence length: 2048
• LoRA Configuration:
  • Rank (r): [Based on Unsloth defaults]
  • Alpha: [Based on Unsloth defaults]
  • Target modules: [Query, Key, Value projections]
• Quantization: 4-bit (QLoRA)

Speeds, Sizes, Times

• Training platform: Google Colab
• Training time: ~60 steps
• Model size: ~3B parameters (base) + LoRA adapters

Technical Specifications

Model Architecture and Objective

• Base Architecture: Llama 3.2 (3B parameters)
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Objective: Supervised Fine-Tuning (SFT) for reflective reasoning

Compute Infrastructure

Hardware

• Google Colab GPU environment

Software

• Framework versions:
  • PEFT: 0.17.1
  • Transformers: Latest compatible version
  • TRL: Latest compatible version
  • Unsloth: Latest version
  • PyTorch: Latest compatible version

Citation

If you use this model, please cite:

@misc{vedai-001-3b,
  author = {Harsh Bopaliya},
  title = {vedai-001-3B: A Reflective Reasoning Model},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/your-username/vedai-001-3B}}
}

Model Card Authors

Harsh Bopaliya

Model Card Contact

For questions or feedback, please open an issue on the model repository.