This is Google's Gemma-4 2B (E2B) instruction-tuned model, fine-tuned on Gemini 3.1 Pro reasoning datasets to improve chain-of-thought reasoning capabilities.

Training Data

Training Configuration

Base Model: google/gemma-4b-it (via unsloth/gemma-4-E2B-it)
Method: LoRA fine-tuning
LoRA Config: r=8, alpha=8, dropout=0.1
Learning Rate: 5e-5
Epochs: 0.5
Framework: Unsloth + TRL

Available Quantizations

Filename	Quant Type	Size	Description
`gemma-4-E2B-Gemini-3.1-Pro-Reasoning-Distill.gguf`	BF16	~5GB	Full precision, best quality
`gemma-4-E2B-Gemini-3.1-Pro-Reasoning-Distill-Q8_0.gguf`	Q8_0	~2.5GB	High quality
`gemma-4-E2B-Gemini-3.1-Pro-Reasoning-Distill-Q5_K_M.gguf`	Q5_K_M	~2GB	Balanced (recommended)
`gemma-4-E2B-Gemini-3.1-Pro-Reasoning-Distill-Q4_K_M.gguf`	Q4_K_M	~1.7GB	Good quality, smaller
`gemma-4-E2B-Gemini-3.1-Pro-Reasoning-Distill-Q3_K_M.gguf`	Q3_K_M	~1.4GB	Smallest

Usage with llama.cpp

bash

# Download a quantized model
wget https://huggingface.co/Ayodele01/gemma-4-E2B-Gemini-3.1-Pro-Reasoning-Distill-GGUF/resolve/main/gemma-4-E2B-Gemini-3.1-Pro-Reasoning-Distill-Q5_K_M.gguf

# Run with llama.cpp
./llama-cli -m gemma-4-E2B-Gemini-3.1-Pro-Reasoning-Distill-Q5_K_M.gguf \
  -p "What is the sum of all prime numbers between 1 and 50?" \
  -n 512

Usage with Ollama

Create a Modelfile:

FROM ./gemma-4-E2B-Gemini-3.1-Pro-Reasoning-Distill-Q5_K_M.gguf

TEMPLATE """<bos><start_of_turn>user
{{ .Prompt }}<end_of_turn>
<start_of_turn>model
"""

PARAMETER stop "<end_of_turn>"
PARAMETER temperature 0.7

Then:

bash

ollama create gemma4-e2b-reasoning -f Modelfile
ollama run gemma4-e2b-reasoning

License

This model is released under the Gemma License.

Ayodele01/gemma-4-E2B-Gemini-3.1-Pro-Reasoning-Distill-GGUF

Gemma-4 E2B Gemini 3.1 Pro Reasoning Distill - GGUF

Model Description