MindSlate: Fine-tuned Gemma-3B for Personal Knowledge Management

Model Description

MindSlate is a fine-tuned version of Google's Gemma-3B model, optimized for personal knowledge management tasks including flashcard generation, reminder processing, content summarization, and task management. The model was trained using Unsloth's efficient fine-tuning techniques for 2x faster training.

• Architecture: Gemma-3B with LoRA adapters
• Model type: Causal Language Model
• Fine-tuning method: 4-bit QLoRA
• Languages: English
• License: Apache 2.0
• Developed by: Srinivas Nampalli
• Finetuned from: unsloth/gemma-3b-E2B-it-unsloth-bnb-4bit

Model Sources

• Repository: https://github.com/Srinivasmec26/MindSlate
• Base Model: unsloth/gemma-3b-E2B-it-unsloth-bnb-4bit

Uses

Direct Use

MindSlate is designed for:

• Automatic flashcard generation from study materials
• Intelligent reminder creation
• Content summarization
• Task extraction and organization
• Personal knowledge base management

Downstream Use

Can be integrated into:

• Educational platforms
• Productivity apps
• Note-taking applications
• Personal AI assistants

Out-of-Scope Use

Not suitable for:

• Medical or legal advice
• High-stakes decision making
• Generating factual content without verification

How to Get Started

from unsloth import FastLanguageModel
import torch

# Load model with Unsloth optimizations
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="Srinivasmec26/MindSlate",
    max_seq_length=2048,
    dtype=torch.float16,
    load_in_4bit=True,
)

# Set chat template
tokenizer = FastLanguageModel.get_chat_template(
    tokenizer,
    chat_template="gemma",  # Use "chatml" or other templates if needed
)

# Create prompt
messages = [
    {"role": "user", "content": "Convert to flashcard: Neural networks are computational models..."},
]

# Generate response
inputs = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt",
).to("cuda")

outputs = model.generate(
    **inputs, 
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.95,
)
print(tokenizer.decode(outputs[0]))

Training Details

Training Data

The model was fine-tuned on a combination of structured datasets:

1. Flashcards Dataset (400 items):

@misc{educational_flashcards_2025,
  title = {Multicultural Educational Flashcards Dataset},
  author = {Srinivas, Yathi Pachauri,  Swarnim Gupta},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Srinivasmec26/Educational-Flashcards-for-Global-Learners}
}

2. Reminders Dataset (100 items):

• Private collection of contextual reminders
• Format: {"input": "Meeting with team", "output": {"time": "2025-08-15 14:00", "location": "Zoom"}}

@misc{educational_flashcards_2025,
  title = {Multicultural Educational Flashcards Dataset},
  author = {Srinivas, Yathi Pachauri,  Swarnim Gupta},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Srinivasmec26/Educational-Flashcards-for-Global-Learners}
}

3. Summaries Dataset (100 items):

• Academic paper abstracts and summaries
• Collected from arXiv and academic publications

@misc{knowledge_summaries_2025,
  title = {Multidisciplinary-Educational-Summaries},
  author = {Srinivas Nampalli, Yathi Pachauri, Swarnim Gupta},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Srinivasmec26/Multidisciplinary-Educational-Summaries}
}

4. Todos Dataset (100 items):

@misc{academic_todos_2025,
   title = {Structured To-Do Lists for Learning and Projects},
  author = {Nampalli Srinivas, Yathi Pachauri, Swarnim Gupta},
  year = {2025},
  publisher = {Hugging Face},
  version   = {1.0},
  url = {https://huggingface.co/datasets/Srinivasmec26/Structured-Todo-Lists-for-Learning-and-Projects}
}

Training Procedure

• Preprocessing: Standardized into ### Input: ... \n### Output: ... format
• Framework: Unsloth 2025.8.1 + Hugging Face TRL
• Hardware: Tesla T4 GPU (16GB VRAM)
• Training Time: 51 minutes for 3 epochs
• LoRA Configuration:

  r=64,           # LoRA rank
  lora_alpha=128, # LoRA scaling factor
  target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                  "gate_proj", "up_proj", "down_proj"],
  

• Optimizer: AdamW 8-bit
• Learning Rate: 2e-4 with linear decay

Evaluation

Comprehensive benchmark results will be uploaded in v1.1. Preliminary metrics:

Metric	Value
Training Loss	0.1284
Perplexity	TBD
Task Accuracy	TBD
Inference Speed	42 tokens/sec (T4)

Technical Specifications

Parameter	Value
Model Size	3B parameters
Quantization	4-bit (bnb)
Max Sequence Length	2048 tokens
Fine-tuned Params	1.66% (91.6M)
Precision	BF16/FP16 mixed
Architecture	Transformer Decoder

Citation

@misc{mindslate2025,
  author = {Srinivas Nampalli },
  title = {MindSlate: Efficient Personal Knowledge Management with Gemma-3B},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Srinivasmec26/MindSlate}},
  note = {Fine-tuned using Unsloth for efficient training}
}

Acknowledgements

• Unsloth for 2x faster fine-tuning
• Google for the Gemma 3n base model
• Hugging Face for TRL library

Model Card Contact

For questions and collaborations:

• Srinivas Nampalli: LinkedIn