InkubaLM-0.4B: Small language model for low-resource African Languages
Model Details
InkubaLM has been trained from scratch using 1.9 billion tokens of data for five African languages, along with English and French data, totaling 2.4 billion tokens of data. Similar to the model architecture used for MobileLLM, we trained this InkubaLM with a parameter size of 0.4 billion and a vocabulary size of 61788. For detailed information on training, benchmarks, and performance, please refer to our full blog post.
Model Description
- Developed by: Lelapa AI - Fundamental Research Team.
- Model type: Small Language Model (SLM) for five African languages built using the architecture design of LLaMA-7B.
- Language(s) (NLP): isiZulu, Yoruba, Swahili, isiXhosa, Hausa, English and French.
- License: CC BY-NC 4.0.
Model Sources
- Repository: TBD
- Paper : TBD
Bias, Risks, and Limitations
The InkubaLM model has been trained on multilingual datasets but does have some limitations. It is capable of understanding and generating content in five African languages: Swahili, Yoruba, Hausa, isiZulu, and isiXhosa, as well as English and French. While it can generate text on various topics, the resulting content may not always be entirely accurate, logically consistent, or free from biases found in the training data. Additionally, the model may sometimes use different languages when generating text. Nonetheless, this model is intended to be a foundational tool to aid research in African languages.
How to Get Started with the Model
Use the code below to get started with the model.
pip install transformers
Running the model on CPU/GPU/multi GPU
- Running the model on CPU
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("lelapa/InkubaLM-0.4B",trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("lelapa/InkubaLM-0.4B",trust_remote_code=True)
text = "Today I planned to"
inputs = tokenizer(text, return_tensors="pt")
input_ids = inputs.input_ids
# Create an attention mask
attention_mask = inputs.attention_mask
# Generate outputs using the attention mask
outputs = model.generate(input_ids, attention_mask=attention_mask, max_length=60,pad_token_id=tokenizer.eos_token_id)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
- Using full precision
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("lelapa/InkubaLM-0.4B", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("lelapa/InkubaLM-0.4B", trust_remote_code=True)
model.to('cuda')
text = "Today i planned to "
input_ids = tokenizer(text, return_tensors="pt").to('cuda').input_ids
outputs = model.generate(input_ids, max_length=1000, repetition_penalty=1.2, pad_token_id=tokenizer.eos_token_id)
print(tokenizer.batch_decode(outputs[:, input_ids.shape[1]:-1])[0].strip())
- Using torch.bfloat16
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
checkpoint = "lelapa/InkubaLM-0.4B"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto",torch_dtype=torch.bfloat16, trust_remote_code=True)
inputs = tokenizer.encode("Today i planned to ", return_tensors="pt").to("cuda")
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))
- Using quantized Versions via bitsandbytes
pip install bitsandbytes accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_8bit=True) # to use 4bit use `load_in_4bit=True` instead
checkpoint = "lelapa/InkubaLM-0.4B"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint, quantization_config=quantization_config, trust_remote_code=True)
inputs = tokenizer.encode("Today i planned to ", return_tensors="pt").to("cuda")
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))
Training Details
Training Data
- For training, we used the Inkuba-mono dataset.
Training Hyperparameters
Hyperparameter | Value |
---|---|
Total Parameters | 0.422B |
Hidden Size | 2048 |
Intermediate Size (MLPs) | 5632 |
Number of Attention Heads | 32 |
Number of Hidden Layers | 8 |
RMSNorm ɛ | 1e^-5 |
Max Seq Length | 2048 |
Vocab Size | 61788 |
Citation
Will be added soon
Model Card Authors
Lelapa AI - Fundamental Research Team
Model Card Contact
- Downloads last month
- 978