🐯 TigerLLM-12B-it

A Bangla instruction-tuned LLM from the TigerLLM family (ACL 2025)

Nishat Raihan, Marcos Zampieri · George Mason University

TigerLLM-9B-it is deprecated. This repository used to be TigerLLM-9B-it. The 9B checkpoints are obsolete and no longer maintained. TigerLLM-12B-it replaces them and scores 5–10% higher than the 9B model on the benchmarks below. Old links to TigerLLM-9B-it redirect here, but please update your code to md-nishat-008/TigerLLM-12B-it.

Model summary

TigerLLM-12B-it is a Bangla instruction-following model built on Gemma 3 (12B). It was fine-tuned with LoRA on Bangla-Instruct, and the LoRA weights are merged into the checkpoint, so you do not need PEFT or a separate base model to run it. It introduces itself as TigerLLM and knows who built it.

Architecture	Gemma3ForConditionalGeneration
Parameters	12.2B
Precision	BF16, 5 safetensors shards (about 22.7 GiB)
Languages	Bangla (primary), English
Fine-tuning data	Bangla-Instruct
License	CC BY 4.0

Only the text decoder was fine-tuned. The vision encoder comes unchanged from Gemma 3 and was not trained or tested on Bangla data.

Benchmarks

Scores on six Bangla benchmarks, from Table 2 of the TigerLLM paper (ACL 2025).

Model	MMLU-bn	PangBench-bn	BanglaQuaD	mHumanEval-bn	BEnQA	BanglaRQA
GPT-3.5	0.55	0.55	0.50	0.56	0.50	0.49
Gemini 1.5 Flash	0.66	0.57	0.62	0.58	0.56	0.61
GPT-4o-mini	0.67	0.62	0.65	0.56	0.60	0.60
LLaMA 3.2 (11B)	0.22	0.19	0.21	0.15	0.18	0.20
Gemma 2 (27B)	0.35	0.51	0.43	0.64	0.50	0.56
Pangea (7B)	0.18	0.15	0.17	0.10	0.14	0.16
Titu-LLM	0.06	0.19	0.08	0.02	0.17	0.21
Bong-LLaMA	0.05	0.12	0.08	0.02	0.15	0.13
Bangla-LLaMA	0.02	0.08	0.05	0.10	0.11	0.09
Bangla-Gemma	0.18	0.15	0.12	0.10	0.22	0.19
TigerLLM 1B, LLaMA 3.2 (deprecated)	0.61	0.55	0.68	0.61	0.59	0.62
TigerLLM-1B-it, Gemma 3	+2–4%	+2–4%	+2–4%	+2–4%	+2–4%	+2–4%
TigerLLM-9B-it (deprecated)	0.72	0.68	0.70	0.63	0.65	0.68
TigerLLM-12B-it	+5–10%	+5–10%	+5–10%	+5–10%	+5–10%	+5–10%

The numeric rows are from the paper. TigerLLM-12B-it scores 5–10% higher than TigerLLM-9B-it, and the current TigerLLM-1B-it scores 2–4% higher than the LLaMA-based TigerLLM 1B.

Quick start

pip install "torch>=2.4" "transformers>=4.57.1" "accelerate>=1.10" sentencepiece safetensors

A GPU with BF16 support and about 24–30 GiB of memory is recommended. For smaller GPUs, see 4-bit loading below.

import torch
from transformers import AutoTokenizer, Gemma3ForConditionalGeneration

MODEL_ID = "md-nishat-008/TigerLLM-12B-it"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = Gemma3ForConditionalGeneration.from_pretrained(
    MODEL_ID,
    dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="sdpa",
).eval()

messages = [{"role": "user", "content": "বাংলায় সহজ ভাষায় ব্যাখ্যা করুন: মেশিন লার্নিং কী এবং এটি কোথায় ব্যবহৃত হয়?"}]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output_ids = model.generate(**inputs, max_new_tokens=256, do_sample=False, repetition_penalty=1.05)

print(tokenizer.decode(output_ids[0, inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

4-bit loading for smaller GPUs

pip install bitsandbytes

import torch
from transformers import BitsAndBytesConfig, Gemma3ForConditionalGeneration

quantization = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

model = Gemma3ForConditionalGeneration.from_pretrained(
    "md-nishat-008/TigerLLM-12B-it",
    device_map="auto",
    quantization_config=quantization,
).eval()

Use the same tokenizer and chat-template code as in the quick start. NF4 cuts memory use and may change output quality slightly.

Prompting tips

• Build prompts with tokenizer.apply_chat_template(...). Do not paste special chat tokens by hand.
• Say the language, task, audience and output format in the prompt, for example: বাংলায় ৮ম শ্রেণির শিক্ষার্থীর উপযোগী করে সালোকসংশ্লেষণ ব্যাখ্যা করুন। উত্তরটি তিনটি ছোট অনুচ্ছেদে দিন।
• Greedy decoding (do_sample=False) can repeat itself on long answers. For long or creative text, try do_sample=True, temperature=0.7, top_p=0.95 or a higher repetition_penalty.
• For multi-turn chat, alternate user and assistant messages and end with the new user turn.

Training

Setting	Value
Base model	Gemma 3 (12B)
Data	Bangla-Instruct: 312,577 unique instruction-response pairs (exact duplicates and one empty row removed), plus 2,000 TigerLLM identity examples in Bangla and English
Method	LoRA on the decoder attention and MLP projections, merged into the weights
LoRA rank / alpha / dropout	32 / 64 / 0
Epochs	1.5 (every example seen at least once)
Max sequence length	2,048 tokens
Optimizer / learning rate	AdamW, 1e-4, cosine schedule
Batch size	about 65K tokens per optimizer step
Precision	BF16
Hardware	1 × NVIDIA RTX PRO 6000 (96 GB), about 15 hours

Training data

Both datasets come from the TigerLLM paper.

• Bangla-Instruct: native Bangla instruction-response pairs. 50 volunteers from Bangladeshi universities wrote 500 seed tasks across 10 categories (culture, academic writing, math, programming, creative writing, science, business, social issues, data analysis, translation). GPT-4 and Claude 3.5 Sonnet expanded them with self-instruct, and every pair passed filters for language, cultural sensitivity, content quality and novelty.
• Bangla-TextBook: 9.9M tokens from 163 open textbooks (Grades 6–12) published by the National Curriculum and Textbook Board of Bangladesh. The paper's 1B and 9B models used it for continual pretraining.

Repository files

File	Contents
model-0000{1..5}-of-00005.safetensors, model.safetensors.index.json	Merged BF16 weights and shard index
config.json, generation_config.json	Model and generation settings
tokenizer.*, added_tokens.json, special_tokens_map.json, chat_template.jinja	Tokenizer and chat format
processor_config.json, preprocessor_config.json	Processor settings

Limitations

• The model can hallucinate and can produce outdated, biased, unsafe or wrong answers. Its knowledge stops at its training data, so recent events may be missing or out of date.
• Quality varies across dialects, romanized Bangla, specialized domains and long inputs.
• Image inputs rely on Gemma 3's original vision encoder and have not been tested in Bangla.
• Do not use it as the only source for medical, legal, financial or other high-stakes decisions. Keep a human in the loop.

License

CC BY 4.0. Check the Gemma 3 terms of use and the dataset licenses before redistribution or commercial use.

Citation

@inproceedings{raihan-zampieri-2025-tigerllm,
    title = "{T}iger{LLM} - A Family of {B}angla Large Language Models",
    author = "Raihan, Nishat  and
      Zampieri, Marcos",
    editor = "Che, Wanxiang  and
      Nabende, Joyce  and
      Shutova, Ekaterina  and
      Pilehvar, Mohammad Taher",
    booktitle = "Proceedings of the 63rd Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 2: Short Papers)",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.acl-short.69/",
    doi = "10.18653/v1/2025.acl-short.69",
    pages = "887--896",
    ISBN = "979-8-89176-252-7"
}

Questions: md.nishat.008@gmail.com