titanthree's picture
Update README.md
a0d89e1 verified
|
Raw History Blame Contribute Delete
3.51 kB
---
language:
- id
- en
license: gemma
library_name: transformers
base_model: google/gemma-3n-E4B-it
tags:
- gemma
- gemma-3n
- unsloth
- lora
- fine-tuned
- indonesian
- english
- text-generation
---
# Orbit-4-Gemma-E4B-it
## Model Description
Orbit-4-Gemma-E4B-it is a language model fine-tuned from Gemma 3n E4B-it, developed to enhance comprehension and knowledge capabilities in the Indonesian language.
The fine-tuning focuses on strengthening the model's ability to follow instructions in Indonesian, understand local context, and generate more natural and relevant responses. The training process utilized a large-scale Indonesian instruction dataset and spanned a full epoch without sampling.
This model retains the multimodal architecture of the Gemma 3n base model, meaning it structurally still supports text, image, audio, and video processing. However, the fine-tuning process was performed exclusively on the text layers, so multimodal performance was not specifically reinforced through this training.
## Model Specifications
| Item | Detail |
|---|---|
| Base model | google/gemma-3n-E4B-it |
| Fine-tuning method | LoRA (r=16, alpha=16) |
| Epoch | 1 full epoch (approx. 29,807 steps) |
| Effective batch size | 4 (batch size 1 x gradient accumulation 4) |
| Learning rate | 2e-4 (linear schedule) |
| Warmup | 3% of total steps |
| Max sequence length | 2048 tokens |
| Optimizer | adamw_8bit |
| Weight decay | 0.01 |
| LoRA dropout | 0 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Merge precision | float32 (performed on CPU) |
| Training hardware | Tesla T4 16GB |
| Framework | Unsloth, TRL, Transformers |
| Prompt format | Gemma 3n chat template |
| Seed | 3407 |
### Prompt Format
The model uses the Gemma 3n chat template as follows:
```text
<bos><start_of_turn>user
{instruction}<end_of_turn>
<start_of_turn>model
{output}<end_of_turn>
```
## Model Capabilities
The model possesses the following capabilities:
* Comprehension of Indonesian-language instructions, including complex and multi-step instructions
* General knowledge and local Indonesian context, covering history, culture, geography, and everyday topics
* Generation of natural Indonesian text for various needs, such as explanations, summaries, and narratives
* Creative writing tasks in Indonesian, such as poetry (*puisi* and *pantun*), and short stories
* Translation between Indonesian and English
* Sentiment analysis and Indonesian text processing
* Basic mathematical reasoning and simple problem-solving
* Inherent multimodal capabilities (text, image, audio, video input; text output), although not reinforced through this fine-tuning
### Limitations
* The model is optimized for Indonesian text; performance on other languages or non-text modalities may decrease compared to the base model
* The model has not been comprehensively tested for multimodal tasks within the Indonesian context
* The model may produce hallucinations; verification of critical facts remains necessary
* The model has not undergone additional safety tuning; deployment in production applications should be accompanied by supervision
* Maximum context is 2048 tokens in accordance with training settings
## Credits
* Base model: Google DeepMind (Gemma 3n)
* Fine-tuning framework: Unsloth
* Dataset: AksaraLLM
* Fine-tuning and training: OpenPathAI
## License
This model follows Google's Gemma license. Commercial use is subject to the Gemma Terms of Use.