🐯 TigerCoder-12B

A Bangla code-generation LLM from the TigerCoder family (LREC 2026)

Nishat Raihan, Antonios Anastasopoulos, Marcos Zampieri · George Mason University

TigerCoder-9B is deprecated. This repository used to be TigerCoder-9B. The 9B checkpoints are obsolete and no longer maintained. TigerCoder-12B replaces them and scores 5–7% higher than the 9B model on the benchmarks below. Old links to TigerCoder-9B redirect here, but please update your code to md-nishat-008/TigerCoder-12B.

Model summary

TigerCoder is a family of code LLMs for Bangla. TigerCoder-12B is the larger member: you describe a programming task in Bangla and it writes the code. It was fine-tuned from TigerLLM-12B-it with LoRA on Bangla-Code-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 TigerCoder and knows who built it.

Architecture	Gemma3ForConditionalGeneration (Gemma 3, 12B)
Parameters	12.2B
Precision	BF16, 5 safetensors shards (about 22.7 GiB)
Base model	TigerLLM-12B-it
Task	Code generation from Bangla prompts
Programming languages	Python, C++, Java, JavaScript
Fine-tuning data	Bangla-Code-Instruct, 300K Bangla instruction-code pairs
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 code or Bangla data.

Benchmarks

Pass@K on Python tasks with Bangla prompts, from the TigerCoder paper (LREC 2026).

Model	mHumanEval P@1	mHumanEval P@10	mHumanEval P@100	MBPP P@1	MBPP P@10	MBPP P@100
GPT-3.5	0.56	0.56	0.59	0.60	0.62	0.62
Gemini-Flash 2.5	0.58	0.61	0.62	0.62	0.62	0.70
Gemma-3 (27B)	0.64	0.65	0.69	0.69	0.70	0.70
TigerLLM (9B) (deprecated)	0.63	0.69	0.72	0.61	0.68	0.73
TigerCoder-1B, paper checkpoint	0.69	0.73	0.77	0.74	0.74	0.81
TigerCoder-1B, current	+2–4%	+2–4%	+2–4%	+2–4%	+2–4%	+2–4%
TigerCoder-9B (deprecated)	0.75	0.80	0.84	0.82	0.84	0.91
TigerCoder-12B (this model)	+5–7%	+5–7%	+5–7%	+5–7%	+5–7%	+5–7%

The numeric rows are from the paper. TigerCoder-12B scores 5–7% higher than TigerCoder-9B, and the current TigerCoder-1B scores 2–4% higher than the paper's TigerCoder-1B.

The paper also tested TigerCoder-9B in other programming languages (Pass@1, Bangla prompts):

Language	mHumanEval P@1	MBPP P@1
Python	0.75	0.82
C++	0.67	0.72
Java	0.62	0.67
JavaScript	0.57	0.62

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/TigerCoder-12B"

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

# "Write a Python function that computes the factorial of a number."
messages = [{"role": "user", "content": "পাইথনে একটি ফাংশন লিখুন যা একটি সংখ্যার ফ্যাক্টরিয়াল গণনা করে।"}]

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=512, 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/TigerCoder-12B",
    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.
• Name the programming language, the function name and signature, and the expected input and output, for example: পাইথনে একটি ফাংশন second_largest(nums) লিখুন যা একটি সংখ্যার তালিকা থেকে দ্বিতীয় বৃহত্তম ভিন্ন সংখ্যাটি ফেরত দেয়।
• Use do_sample=False for Pass@1-style answers. For several candidate solutions (Pass@10, Pass@100), sample with do_sample=True, temperature=0.7, top_p=0.95.
• You can also paste code and ask, in Bangla, for an explanation, a bug fix or a translation to another programming language.

Training

Setting	Value
Base model	TigerLLM-12B-it
Data	Bangla-Code-Instruct: 294,883 unique instruction-code pairs (exact duplicates and one empty row removed), plus 2,000 TigerCoder 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), 3,047 optimizer steps
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 setup the paper reports for TigerCoder-9B

Hyperparameter	Value
Training data	Bangla-Code-Instruct (300K examples)
Max sequence length	2048
Batch size (train / eval)	32
Gradient accumulation steps	8
Epochs	3
Learning rate	1 × 10⁻⁶
Weight decay	0.04
Warm-up	15% of steps
Optimizer	AdamW
LR scheduler	Cosine
Precision	BF16
Hardware	NVIDIA A100 (40 GB)

Training data

Bangla-Code-Instruct has three 100K subsets:

Subset	Size	How it was built	Prompt origin	Code origin
SI (Self-Instruct)	100K	5,000 expert seeds expanded with GPT-4o	Semi-natural	Synthetic
Syn (Synthetic)	100K	Generated with GPT-4o and Claude 3.5 Sonnet	Synthetic	Synthetic
TE (Translated)	100K	Evol-Instruct machine-translated with NLLB-200	Translated	Natural

All SI and Syn code passed a syntax check (ast.parse) and an execution test (Python 3.13.0, 10 s timeout, 16 GB memory).

Findings from the paper

1. Bangla prompts hurt code LLMs. Most models score 20 to 50+ percentage points lower on Bangla coding prompts than on the same prompts in English.
2. Translating the prompt to English does not fix it. Machine translation garbles programming terms: "অক্ষর" (character) becomes "letter", "চলক" (variable) becomes "clever", and "স্ট্রিং" (string) becomes "rope".
3. Targeted data beats scale. TigerCoder-1B beats models 27 times its size, and TigerCoder-9B led the strongest prior baselines by 11–18%.

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

• TigerCoder is trained for Bangla code generation. On general language tasks, use TigerLLM-12B-it.
• The training data is synthetic or machine-translated, and may carry the biases and artifacts of the models that produced it.
• Evaluation covers MBPP-Bangla and mHumanEval-Bangla only. Production-scale coding tasks have not been benchmarked.
• Run and review generated code before you use it.

License

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

Citation

@inproceedings{raihan-etal-2026-tigercoder,
    title = "{T}iger{C}oder: A Novel Suite of {LLM}s for Code Generation in {B}angla",
    author = "Raihan, Nishat  and
      Anastasopoulos, Antonios  and
      Zampieri, Marcos",
    editor = "Piperidis, Stelios  and
      Bel, N{\'u}ria  and
      van den Heuvel, Henk  and
      Ide, Nancy  and
      Krek, Simon  and
      Toral, Antonio",
    booktitle = "Proceedings of the Fifteenth Language Resources and Evaluation Conference",
    month = may,
    year = "2026",
    address = "Palma de Mallorca, Spain",
    publisher = "ELRA Language Resource Association",
    url = "https://aclanthology.org/2026.lrec-1.238/",
    doi = "10.63317/5nampb63np3m",
    pages = "3044--3054"
}

Related work

@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"
}

@inproceedings{raihan-etal-2025-mhumaneval,
    title = "m{H}uman{E}val - A Multilingual Benchmark to Evaluate Large Language Models for Code Generation",
    author = "Raihan, Nishat  and
      Anastasopoulos, Antonios  and
      Zampieri, Marcos",
    editor = "Chiruzzo, Luis  and
      Ritter, Alan  and
      Wang, Lu",
    booktitle = "Proceedings of the 2025 Conference of the Nations of the {A}mericas Chapter of the {A}ssociation for {C}omputational {L}inguistics: Human Language Technologies (Volume 1: Long Papers)",
    month = apr,
    year = "2025",
    address = "Albuquerque, New Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.naacl-long.570/",
    doi = "10.18653/v1/2025.naacl-long.570",
    pages = "11432--11461",
    ISBN = "979-8-89176-189-6"
}

Questions: mraihan2@gmu.edu