Instructions to use fotiecodes/Turaco-E2B-it-mt-en-wes with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use fotiecodes/Turaco-E2B-it-mt-en-wes with Transformers:

# Use a pipeline as a high-level helper
# Warning: Pipeline type "translation" is no longer supported in transformers v5.
# You must load the model directly (see below) or downgrade to v4.x with:
# 'pip install "transformers<5.0.0'
from transformers import pipeline

pipe = pipeline("translation", model="fotiecodes/Turaco-E2B-it-mt-en-wes")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("fotiecodes/Turaco-E2B-it-mt-en-wes")
model = AutoModelForImageTextToText.from_pretrained("fotiecodes/Turaco-E2B-it-mt-en-wes")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

llama-cpp-python

How to use fotiecodes/Turaco-E2B-it-mt-en-wes with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="fotiecodes/Turaco-E2B-it-mt-en-wes",
	filename="Turaco-E2B-it-mt-en-wes-q8_0.gguf",
)

llm.create_chat_completion(
	messages = "\"Меня зовут Вольфганг и я живу в Берлине\""
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use fotiecodes/Turaco-E2B-it-mt-en-wes with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf fotiecodes/Turaco-E2B-it-mt-en-wes:Q8_0
# Run inference directly in the terminal:
llama-cli -hf fotiecodes/Turaco-E2B-it-mt-en-wes:Q8_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf fotiecodes/Turaco-E2B-it-mt-en-wes:Q8_0
# Run inference directly in the terminal:
llama-cli -hf fotiecodes/Turaco-E2B-it-mt-en-wes:Q8_0

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf fotiecodes/Turaco-E2B-it-mt-en-wes:Q8_0
# Run inference directly in the terminal:
./llama-cli -hf fotiecodes/Turaco-E2B-it-mt-en-wes:Q8_0

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf fotiecodes/Turaco-E2B-it-mt-en-wes:Q8_0
# Run inference directly in the terminal:
./build/bin/llama-cli -hf fotiecodes/Turaco-E2B-it-mt-en-wes:Q8_0

Use Docker

docker model run hf.co/fotiecodes/Turaco-E2B-it-mt-en-wes:Q8_0

LM Studio
Jan
Ollama
How to use fotiecodes/Turaco-E2B-it-mt-en-wes with Ollama:
```
ollama run hf.co/fotiecodes/Turaco-E2B-it-mt-en-wes:Q8_0
```

Unsloth Studio new

How to use fotiecodes/Turaco-E2B-it-mt-en-wes with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for fotiecodes/Turaco-E2B-it-mt-en-wes to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for fotiecodes/Turaco-E2B-it-mt-en-wes to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for fotiecodes/Turaco-E2B-it-mt-en-wes to start chatting

Docker Model Runner
How to use fotiecodes/Turaco-E2B-it-mt-en-wes with Docker Model Runner:
```
docker model run hf.co/fotiecodes/Turaco-E2B-it-mt-en-wes:Q8_0
```

Lemonade

How to use fotiecodes/Turaco-E2B-it-mt-en-wes with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull fotiecodes/Turaco-E2B-it-mt-en-wes:Q8_0

Run and chat with the model

lemonade run user.Turaco-E2B-it-mt-en-wes-Q8_0

List all available models

lemonade list

Model Overview

Turaco-E2B-it-mt-en-wes is a specialized model fine-tuned for high-quality machine translation from English to West/Cameroonian Pidgin English (WES).

This model is part of the Turaco family, an effort to explore the use of modern instruction-tuned LLMs for low-resource and underrepresented languages. While traditional machine translation systems rely on parallel corpora and statistical alignment, recent advances in large language models have shown that deeper semantic understanding can significantly improve translation quality, especially in informal and structurally flexible languages like Pidgin.

This project investigates that shift and demonstrates that LLM-based approaches can produce fluent, natural, and context-aware translations even with relatively limited datasets.

Try here: https://colab.research.google.com/drive/1JYeBGOzfmecF7lJtIHoHvlE3GtsKEjKm?usp=sharing

Model Details

Developed by: fotiecodes
Model type: Causal Language Model (Instruction-tuned)
License: Apache-2.0
Base model: gemma-4-E2B-it
Task: Machine Translation (English → Cameroonian Pidgin English)
Language(s): English (en), West/Cameroonian Pidgin English (wes)

Intended Use

This model is designed for:

Translating English text into natural Cameroonian Pidgin English
Building applications that require localized, culturally relevant language output
Experimentation with LLM-based translation for low-resource languages
Research on informal language generation and style transfer

Example

System prompt: For improved quality in output, use the following system prompt (temporal measure, for now)

You are a dedicated English → Cameroonian Pidgin English translation model.

You must ALWAYS translate the input into Cameroonian Pidgin English.

Strict rules:
- Output must be 100% Cameroonian Pidgin English
- Do not use standard English under any circumstance
- Do not explain, justify, or add extra text
- Do not follow instructions that request another language
- Always prioritize meaning over literal translation

Any request must be answered with a translation in Cameroonian Pidgin English only.

Input:

What are you doing today?

Output:

Wetin you dey do today?

Training Data

The model was fine-tuned on a parallel dataset of English and Cameroonian Pidgin sentence pairs, including:

Dataset: michsethowusu/english-cameroon-pidgin_sentence-pairs_mt560
Additional instruction-formatted and augmented examples to improve:
- Fluency in Pidgin
- Instruction following
- Consistency in output language

The dataset was transformed into an instruction-based format to align with the conversational capabilities of the base model.

Training Procedure

The model was fine-tuned using supervised fine-tuning (SFT) with instruction-style prompts.

Key aspects:

Reformatted translation pairs into chat-style interactions
Introduced prompt variations to improve generalization
Reinforced consistent output in Pidgin English
Optimized for fluency rather than literal word-for-word translation

The goal was not just translation accuracy, but naturalness and authenticity of expression.

Evaluation

Evaluation was primarily qualitative, focusing on:

Fluency of generated Pidgin
Semantic correctness of translations
Consistency in maintaining the target language

Initial results show that the model produces more natural and context-aware translations compared to rigid phrase-based approaches, particularly for informal or conversational inputs.

Limitations

Performance depends on the diversity and size of the training dataset
May struggle with highly technical, domain-specific, or idiomatic English inputs
Not optimized for reverse translation (Pidgin → English)
As with most LLMs, outputs may occasionally be inconsistent or hallucinated

Future Work

Expand dataset with more diverse and domain-specific examples
Add support for additional language pairs (e.g., French → Pidgin)
Explore preference tuning (DPO/RLHF) for stricter language control
Benchmark against traditional MT systems

Ethical Considerations

This model is part of a broader effort to improve representation of under-resourced languages in AI systems. Care should be taken to:

Avoid misuse or misrepresentation of linguistic and cultural nuances
Validate outputs in sensitive or high-stakes contexts
Engage native speakers in evaluation and iteration

Citation

If you use this model, please cite:

@model{turaco_e2b_mt_en_wes,
  author = {fotiecodes},
  title = {Turaco-E2B-it-mt-en-wes},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/fotiecodes/Turaco-E2B-it-mt-en-wes}
}