sartifyllc/Pawa-mini-V0.1

PAWA: Swahili SML for Various Tasks

---

Overview

PAWA is a Swahili-specialized language model designed to excel in tasks requiring nuanced understanding and interaction in Swahili and English. It leverages supervised fine-tuning (SFT) and Direct Preference Optimization (DPO) for improved performance and consistency. Below are the detailed model specifications, installation steps, usage examples, and its intended applications.

---

Model Details

• Model Name: Pawa-mini-V0.1
• Model Type: PAWA
• Architecture:
  • 2B Parameter Gemma-2 Base Model
  • Enhanced with Swahili SFT and DPO datasets.
• Languages Supported:
  • Swahili
  • English
  • Custom tokenizer for multi-language flexibility.
• Primary Use Cases:
  • Contextually rich Swahili-focused tasks.
  • General assistance and chat-based interactions.
• License: Custom/Contact Author for terms of use.

---

Installation and Setup

Ensure the necessary libraries are installed and up-to-date:

!pip uninstall transformers -y && pip install --upgrade --no-cache-dir "git+https://github.com/huggingface/transformers.git"
!pip uninstall unsloth -y && pip install --upgrade --no-cache-dir "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install datasets

---

Model Loading

You can load the model using the following code snippet:

from unsloth import FastLanguageModel
import torch

model_name = "sartifyllc/Pawa-mini-V0.1"
max_seq_length = 2048  
dtype = None  
load_in_4bit = False  

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_name,
    max_seq_length=max_seq_length,
    dtype=dtype,
    load_in_4bit=load_in_4bit,
)

---

Chat Template Configuration

For a seamless conversational experience, configure the tokenizer with the appropriate chat template:

from unsloth.chat_templates import get_chat_template
FastLanguageModel.for_inference(model) # Enable native 2x faster inference

tokenizer = get_chat_template(
    tokenizer,
    chat_template="chatml",  # Supports templates like zephyr, chatml, mistral, etc.
    mapping={"role": "from", "content": "value", "user": "human", "assistant": "gpt"},  # ShareGPT style
    map_eos_token=True,  # Maps <|im_end|> to </s>
)

---

Usage Example

Generate a short story in Swahili:

messages = [{"from": "human", "value": "Tengeneza hadithi fupi"}]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
).to("cuda")

from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer)
_ = model.generate(input_ids=inputs, streamer=text_streamer, max_new_tokens=128, use_cache=True)

---

Training and Fine-Tuning Details

• Base Model: Gemma-2-2B
• Continue Pre-Training: 3B Swahili Tokens
• Fine-tuning: Enhanced with Swahili SFT datasets for improved contextual understanding.
• Optimization: Includes DPO for deterministic and consistent responses.

---

Intended Use Cases

• General Assistance:
  Provides structured answers for general-purpose use.

• Interactive Q&A:
  Designed for general-purpose chat environments.

• RAG (Retrieval-Augmented Generation):
  Works best for RAG and specific use cases.

---

Limitations

• Biases:
  The model may exhibit biases inherent in its fine-tuning datasets.

• Generalization:
  May struggle with tasks outside the trained domain.

• Hardware Requirements:

  • Optimal performance requires GPUs with high memory (e.g., Tesla V100 or T4).
  • Supports 4-bit quantization for reduced memory usage.

Feel free to reach out for further guidance or collaboration opportunities regarding PAWA!