LilMoo-v0.2

Model Summary

LilMoo-v0.2 is a decoder-transformer natively pretrained in Hindi and English. LilMoo is part of the Polygl0t initiative to advance language models for low-resource languages.

Details

• Architecture: a Transformer-based model (llama)
• Size: 670,127,616 parameters
• Context length: 4096 tokens
• Dataset(s):
  • - Polygl0t/gigalekh-v1
  • allenai/big-reasoning-traces
  • HuggingFaceTB/smollm-corpus
  • HuggingFaceTB/finemath
  • HuggingFaceFW/fineweb-edu
  • allenai/math-meta-reasoning-filtered
  • nvidia/OpenScience
• Language(s): Hindi, English
• Batch size: 2,097,152 tokens
• Number of steps: 170,000
• GPU: 8 NVIDIA A100-SXM4-80GB
• Training time: ~ 327.50 hours
• Emissions: 513.05 KgCO2eq (Germany)
• Total energy consumption: 1346.76 kWh

This repository has the source code used to train this model. The full configuration used for training is available in the following config files:

• Stage 1 (warmup + stable): config_stage_1.yaml
• Stage 2 (stable): config_stage_2.yaml
• Stage 3 (stable + decay): config_stage_3.yaml

Learning Rate

This model was trained with a linear learning rate (i.e., warmup-stable-decay). Warmup was done over the first 2,000 steps to a peak learning rate of 7e-4, followed by a stable phase until step 149,000 (this revision is named end-of-stable-ckpt), and then a negative square-root decay to 0 for the last 21,000 steps. Checkpoints were saved every 5,000 steps, which equates to approximately 10 billion tokens. The end-of-stable-ckpt can be used for continuing training with a stable learning rate if desired.

The main branch of this repository contains the final checkpoint saved at step 170,000. All other checkpoints are available as separate branches. To load a specific checkpoint, you can use the following code snippet:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Polygl0t/LilMoo-v0.2"
revision = "stage1-step-05000"  # Change this to the desired checkpoint branch
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, revision=revision)

Or, you can access all the revisions for the models via the following code snippet:

from huggingface_hub import list_repo_refs
out = list_repo_refs("Polygl0t/LilMoo-v0.2")
branches = [b.name for b in out.branches]
print(branches)

Learning Curves Gradient Norm (L2)

Intended Uses

The primary intended use LilMoo is to serve as foundations for research and development involving native Hindi language modeling. Checkpoints saved during training are designed to provide a controlled setting for performing comparative experiments, specifically regarding the effects of active pretraining on the performance of currently available benchmarks. You may also fine-tune and adapt LilMoo for deployment if your use follows the Apache 2.0 license. If you decide to use LilMoo as a basis for your fine-tuned model, please conduct your own risk and bias assessment.

Out-of-scope Use

• LilMoo is not intended for deployment. It is not an out-of-the-box product and should not be used for human-facing interactions.

• LilMoo is for the Hindi language only and is unsuitable for text generation tasks in other languages.

• LilMoo has not been fine-tuned for downstream tasks.

Basic usage

from transformers import GenerationConfig, TextGenerationPipeline, AutoTokenizer, AutoModelForCausalLM
import torch

# Specify the model and tokenizer
model_id = "Polygl0t/LilMoo-v0.2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

# Specify the generation parameters as you like
generation_config = GenerationConfig(
    **{
    "do_sample": True,
    "max_new_tokens": 150,
    "renormalize_logits": True,
    "repetition_penalty": 1.2,
    "temperature": 0.1,
    "top_k": 50,
    "top_p": 1.0,
    "use_cache": True, 
  }
)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
generator = TextGenerationPipeline(model=model, task="text-generation", tokenizer=tokenizer, device=device)

# Generate text
prompt = "भारत की राजधानी क्या है?"
completion = generator(prompt, generation_config=generation_config)
print(completion[0]['generated_text'])

Limitations

Like almost all other language models trained on large text datasets scraped from the web, the LilMoo shows behavior that does not make it an out-of-the-box solution to many real-world applications, especially those requiring factual, reliable, and nontoxic text generation. LilMoo is subject to the following:

• Hallucinations: LilMoo can produce content that can be mistaken as true facts, but are misleading or entirely false, i.e., hallucination.

• Biases and Toxicity: LilMoo inherits the social and historical stereotypes from the data used to train it. Given these biases, the model can produce toxic content, i.e., harmful, offensive, or detrimental to individuals, groups, or communities.

• Language Limitations: LilMoo is primarily designed to interact with Hindi. Other languages might challenge its comprehension, leading to potential misinterpretations or errors in response.

• Repetition and Verbosity: LilMoo may get stuck on repetition loops (especially if the repetition penalty during generations is set to a meager value) or produce verbose responses unrelated to the prompt it was given.

Hence, even though LilMoo is released with a permissive license, we urge users to perform their risk analysis on them if they intend to use them for real-world applications.

Evaluations

The table below compares our two versions of LilMoo against other base models of similar size. The NPM (Normalized Performance Metric) is a metric designed to provide a balanced view of model performance across various tasks, accounting for the inherent difficulty of each task. It normalizes the performance of each model on a given task by comparing it to a baseline performance, which represents a the performance of a random model.

Note: In the table and plots below, benchmark scores are reported as normalized accuracy, and NPM is calculated from accuracy.

🏆 HellaSwag 🏆 ARC 🏆 MMLU 🏆 MILU 🏆 CSQA 🏆 Global PIQA Aggregate NPM Across Benchmarks Performance vs Compute

This plot compares the compute requirements (measured as C = 6 * N * D, where N is the number of parameters and D is the number of tokens processed) against the performance of each model (measured by NPM mean). It highlights the trade-offs between model size, training data, and performance. LilMoo models are compared against two baseline models: Qwen2.5-0.5B and Qwen3-0.6B-Base, which are state-of-the-art multilingual models.

Text Generation Examples

All models were prompted with the same Hindi prompts, and the generations were produced with the default parameters specified in the generation_config.json of each model. All models were set to generate up to 250 new tokens. LilMoo models generally produce more text given their custom tokenizer and vocabulary, which is more adapted to Hindi.

Cite as 🤗

@misc{shiza2026lilmoo,
      title={{Raising Bars, Not Parameters: LilMoo Compact Language Model for Hindi}}, 
      author={Shiza Fatimah and Aniket Sen and Sophia Falk and Florian Mai and Lucie Flek and Nicholas Kluge Corr{\^e}a},
      year={2026},
      eprint={2603.03508},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.03508}, 
}

Aknowlegments

Polyglot is a project funded by the Federal Ministry of Education and Research (BMBF) and the Ministry of Culture and Science of the State of North Rhine-Westphalia (MWK) as part of TRA Sustainable Futures (University of Bonn) and the Excellence Strategy of the federal and state governments.

We also gratefully acknowledge the granted access to the Marvin cluster hosted by University of Bonn along with the support provided by its High Performance Computing & Analytics Lab.

License

LilMoo is licensed under the Apache License, Version 2.0. For more details, see the LICENSE file.