Ila (Indic GPT)

Ila is a compact Indic-language causal LLM trained for generation in the Hindi language. Due to its size (160M parameters), it is extremely lightweight and can run on most laptop GPUs whilst performing decent generation for short-form content.

---

Model name: Ila (Indic GPT v3.0.3)

Estimated parameters: ~160M

Architecture: LlamaForCausalLM transformer (decoder only)

Language(s): Hindi (hi)

Intended use: text generation, completion, short-form creative writing, assistive language tools.

---

Model configuration

• hidden_size: 768
• intermediate_size: 3072
• num_hidden_layers: 10
• num_attention_heads: 12
• max_position_embeddings: 512
• vocab_size: 64014
• rms_norm_eps: 1e-6
• tie_word_embeddings: true
• torch_dtype: float16

Training data

• Source: ai4bharat/sangraha(https://huggingface.co/datasets/ai4bharat/sangraha) (verified/hin data splits used for training).
• Rows: ~1.2 million rows of cleaned Hindi text. Tokenized with the ai4bharat/IndicBART tokenizer, truncation to 512 subword tokens.
• Training: trained for 8,000 steps on an RTX 3050 (4GB VRAM)

Usage

Install requirements:

pip install -r requirements.txt
pip install transformers datasets accelerate safetensors

Usage (Hugging Face Hub)

You can load it directly from HuggingFace :

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("DireDreadlord/indic-gpt-v3")
model = AutoModelForCausalLM.from_pretrained("DireDreadlord/indic-gpt-v3")

prompt = "मुलायम यादव ने"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
 **inputs,
 max_new_tokens=64,
 do_sample=True,
 top_k=50,
 top_p=0.95,
 temperature=0.7,
 pad_token_id=tokenizer.eos_token_id,
 eos_token_id=tokenizer.eos_token_id,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Generation tips and parameters

• For creative generation use do_sample=True, top_p=0.9, temperature=0.7.
• For deterministic completions use do_sample=False, num_beams=4.
• The model supports up to 512 input + generated tokens but memory will constrain this.

Limitations & Safety

• Trained on scraped/curated Hindi text; it may reproduce social biases present in training data.
• Not safe for high-stakes advice; may hallucinate facts.
• Users should employ content filters for downstream deployments and include user-facing disclaimers where appropriate.

License

cc-by-4.0