---
base_model: unsloth/llama-3.2-3b-instruct-bnb-4bit
tags:
- text-generation-inference
- transformers
- unsloth
- llama
- trl
license: apache-2.0
language:
- en
- sw
datasets:
- wikimedia/wikipedia
- Mollel/alpaca-swahili
- Mollel/swahili_pretrain_data
library_name: peft
---

# Model description

The model can be used for Swahili language generation, translation, and other NLP tasks, especially focused on the pretraining and fine-tuning domains.
It has been pre-trained and fine-tuned specifically for Swahili language tasks with the Unsloth framework. 

This is a development version and it's not recommended for general use.

- **Developed by:** calcpy
- **License:** apache-2.0
- **Finetuned from model :** unsloth/llama-3.2-3b-instruct-bnb-4bit

This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.

[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)


## Out-of-Scope Use

The model is not designed for tasks outside of the Swahili language 
or tasks requiring highly factual precision in domains not covered by the training datasets.


## Bias, Risks, and Limitations

The model inherits any potential biases present in the Swahili Wikipedia and Mollel's dataset. Users should be cautious when applying this model to sensitive applications.

### Recommendations

Users should perform bias evaluations specific to their use case and ensure that any downstream applications consider potential ethical implications.

## How to Get Started with the Model

Use the code below to get started with the model.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load the model and tokenizer
model = AutoModelForCausalLM.from_pretrained("path_to_your_model")
tokenizer = AutoTokenizer.from_pretrained("path_to_your_model")

# Example inference
instruction = "Endelea mlolongo wa fibonacci:"
input_data = "1, 1, 2, 3, 5, 8,"
prompt = f"Chini ni maagizo ambayo yanaelezea kazi. Andika jibu ambalo linakamilisha ombi ipasavyo.\n### Maagizo:\n{instruction}\n\n{input_data}\n### Jibu:\n"

inputs = tokenizer([f"{prompt}"], return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=64, use_cache=True)
print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
```

In this example, the model generates the continuation of the Fibonacci sequence in Swahili.


## Training Hyperparameters

- ** Training regime: Mixed precision (fp16/bf16)
- ** Batch size: 2 per device
- ** Max steps: 24,000 for pretraining, 1,200 for fine-tuning
- ** Learning rate: 5e-5 (1e-5 for embeddings)
- ** Warmup steps: 100 for pretraining, 10 for fine-tuning
- ** Weight decay: 0.01 (pretraining), 0.00 (fine-tuning)

## Evaluation

The model was only manually evaluated on the Alpaca Swahili dataset for instruction-following capabilities.

## Metrics

Evaluation metrics will be required for language generation quality and instruction-following precision.


## Summary

This is a technical release of a small test model to test pre-training and fine-tuning on a single GPU. 


## Compute Infrastructure

- **OS** Ubuntu 22.04.5 LTS
- **Hardware Type:** NVIDIA GeForce RTX 4090 24 GiB
- **Hours used:** ~12 hours