Krvn – Gemma-3N (4B) Fine-tuned on FineTome-100k
Developed by: gawadx1
License: Apache-2.0
Base Model: unsloth/gemma-3n-e4b-it-unsloth-bnb-4bit
📌 Overview
Krvn is a fine-tuned variant of the Gemma-3N (4B) language model, optimized using the Unsloth accelerated training framework and the Hugging Face TRL library.
The model is trained on the FineTome-100k dataset to improve instruction quality, reasoning depth, and conversational consistency.
With Unsloth optimizations, Krvn trains approximately 2× faster than traditional Transformer fine-tuning pipelines, making it accessible for developers running on consumer GPUs or cloud notebooks.
Krvn is designed as a lightweight yet capable assistant, making it ideal for real-world applications that require fast inference and low hardware requirements.
🧩 Key Features
- 🔹 Optimized for instruction following
- 🔹 Enhanced reasoning and coherence
- 🔹 4-bit quantized, enabling efficient local inference
- 🔹 Compatible with Transformers, Text Generation Inference, and LLM runners
- 🔹 Fast fine-tuning using Unsloth
- 🔹 Based on an aligned Gemma-3N model for safer responses
📚 Training Data
Dataset: mlabonne/FineTome-100k
A curated, high-quality instruction dataset containing:
- Multi-turn conversations
- Thought-through responses
- Problem-solving and reasoning tasks
- Knowledge-rich instructions
- Chain-of-thought examples
- General-purpose conversational prompts
FineTome-100k is widely used for improving:
- Assistant alignment
- Instruction clarity
- Relevance and coherence
- Reasoning performance
⚙️ Training Configuration
- Framework: Unsloth (accelerated fine-tuning)
- Libraries: TRL + Transformers
- Precision: 4-bit QLoRA / bnb quantization
- Objective: Supervised Fine-Tuning (Instruction Tuning)
- Training Speed: ~2× faster vs. standard HF training
- Hardware: (optional—add GPU details if you want)
If you'd like, I can generate a full JSON training config block.
🚀 Intended Use
Krvn is suitable for:
- AI assistants
- Chatbots
- Reasoning and explanation tasks
- Educational tools
- Personal AI agents
- Low-resource or on-device inference
- Research in model alignment and instruction design
❌ Not Intended For
- High-risk decision making (medical, legal, financial, etc.)
- Real-time safety-critical systems
- Producing verified factual content without supervision
- Applications requiring guaranteed correctness
📥 Usage Example
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "gawadx1/Krvn"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
prompt = "Explain why data augmentation helps improve model generalization."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))