Edit model card

Ma-layala-mba

Welcome to Ma-layala-mba, a base Indic language model designed to push the boundaries of NLP for Indian languages. It is based on the Mamba series of state space models.

Thumbnail

Model Description

Ma-layala-mba is a state-of-the-art S6 SSM model specifically crafted for the South Indian regional and state language of Kerala: Malayalam. It integrates traditional Attention mechanisms with innovative approaches such as MLPs and State Space Models (SSMs) to handle complex linguistic features and achieve high accuracy in language understanding and generation.

  • Model Type: A 128M Jamba model finetuned on ~1 million samples of Malayalam prompt-response pairs from a subset of the IndicCorp Dataset
  • Language(s): Malayalam
  • License: GNU General Public License v3.0
  • Training Precision: bfloat16

Example Usage

Here's a quick example to get you started with the Ma-layala-mba model:

from transformers import MaLayalaMbaForCausalLM, AutoTokenizer, pipeline

model = MaLayalaMbaForCausalLM.from_pretrained(
    "aoxo/Ma-layala-mba_Tiny_128M",
    # load_in_8bit=True, # Set this depending on the GPU you have
    torch_dtype=torch.bfloat16,
    device_map={"": 0}, # Set this depending on the number of GPUs you have
    local_files_only=False # Optional
)
model.eval()

tokenizer = AutoTokenizer.from_pretrained("aoxo/Ma-layala-mba_Tiny_128M")

input_ids = tokenizer("മലയാളം പര്യായപദങ്ങളിൽ ഒരു പരീക്ഷ പേപ്പർ ഉണ്ടാക്കുക", return_tensors='pt').to(model.device)["input_ids"]

outputs = model.generate(input_ids, max_new_tokens=100)

print(tokenizer.batch_decode(outputs))

Example Output:

മലയാളം പര്യായപദങ്ങളിൽ ഒരു പരീക്ഷ പേപ്പർ ഉണ്ടാക്കുക

a. വലിയ - __________
b. രസം - __________
c. സുഖം - __________
d. പ്രകാശം - __________
e. വേഗം - __________

Usage Note

Please be aware that this model has not undergone comprehensive detoxification or censorship. While it exhibits strong linguistic capabilities, there is a possibility of generating content that may be deemed harmful or offensive. We advise users to apply discretion and closely monitor the model's outputs, especially in public or sensitive settings.

Meet the Developers

Downloads last month
23
Safetensors
Model size
128M params
Tensor type
BF16
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Dataset used to train aoxo/Ma-layala-mba_Tiny_128M