91 lines
1.9 KiB
Markdown
91 lines
1.9 KiB
Markdown
|
---
|
||
|
|
license: gemma
|
||
|
|
base_model: google/gemma-3-1b-it
|
||
|
|
tags:
|
||
|
|
- fine-tuned
|
||
|
|
- gguf
|
||
|
|
- gemma
|
||
|
|
- instruct
|
||
|
|
- 1b
|
||
|
|
model_type: gemma3
|
||
|
|
quantized: f16
|
||
|
|
datasets:
|
||
|
|
- custom
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
---
|
||
|
|
|
||
|
|
# Fine-tuned Gemma 3 1B IT (GGUF)
|
||
|
|
|
||
|
|
This is a fine-tuned version of Google's Gemma 3 1B IT model, converted to GGUF format for efficient inference.
|
||
|
|
|
||
|
|
## Model Details
|
||
|
|
|
||
|
|
- **Base Model**: google/gemma-3-1b-it
|
||
|
|
- **Fine-tuning Method**: QLoRA (Quantized Low-Rank Adaptation)
|
||
|
|
- **Format**: GGUF (F16 precision)
|
||
|
|
- **Size**: ~1.9GB
|
||
|
|
|
||
|
|
## Training Details
|
||
|
|
|
||
|
|
- **Adapter**: QLoRA with rank 32, alpha 16
|
||
|
|
- **Target Modules**: up_proj, down_proj, gate_proj, q_proj, k_proj, v_proj, o_proj
|
||
|
|
- **Sequence Length**: 2048
|
||
|
|
- **Training Data**: Custom dataset with sample packing
|
||
|
|
- **Epochs**: 3
|
||
|
|
- **Learning Rate**: 0.0004 with cosine scheduler
|
||
|
|
- **Optimizer**: AdamW BNB 8-bit
|
||
|
|
|
||
|
|
## Usage
|
||
|
|
|
||
|
|
This GGUF model can be used with various inference engines:
|
||
|
|
|
||
|
|
### llama.cpp
|
||
|
|
```bash
|
||
|
|
./llama-server -m model.gguf
|
||
|
|
```
|
||
|
|
|
||
|
|
### Ollama
|
||
|
|
```bash
|
||
|
|
# Create a Modelfile
|
||
|
|
FROM model.gguf
|
||
|
|
TEMPLATE """<start_of_turn>user
|
||
|
|
{{ .Prompt }}<end_of_turn>
|
||
|
|
<start_of_turn>model
|
||
|
|
"""
|
||
|
|
|
||
|
|
# Import the model
|
||
|
|
ollama create my-gemma-model -f Modelfile
|
||
|
|
ollama run my-gemma-model
|
||
|
|
```
|
||
|
|
|
||
|
|
### Python with llama-cpp-python
|
||
|
|
```python
|
||
|
|
from llama_cpp import Llama
|
||
|
|
|
||
|
|
llm = Llama(model_path="model.gguf")
|
||
|
|
output = llm("Tell me about artificial intelligence", max_tokens=512)
|
||
|
|
print(output)
|
||
|
|
```
|
||
|
|
|
||
|
|
## Chat Template
|
||
|
|
|
||
|
|
This model uses Gemma 3's chat template:
|
||
|
|
```
|
||
|
|
<start_of_turn>user
|
||
|
|
{user_message}<end_of_turn>
|
||
|
|
<start_of_turn>model
|
||
|
|
{assistant_response}<end_of_turn>
|
||
|
|
```
|
||
|
|
|
||
|
|
## Limitations
|
||
|
|
|
||
|
|
- This is a small 1B parameter model with inherent limitations
|
||
|
|
- Fine-tuned for specific use cases - performance may vary on other tasks
|
||
|
|
- GGUF conversion may introduce minor numerical differences compared to the original model
|
||
|
|
|
||
|
|
## License
|
||
|
|
|
||
|
|
This model inherits the license from the base Gemma model. Please refer to Google's Gemma license for usage terms.
|