README.md

---
license: cc-by-4.0
language:
  - bn
  - en
library_name: transformers
pipeline_tag: text-generation
tags:
  - code
  - bangla
  - bengali
  - code-generation
  - nlp
  - low-resource
datasets:
  - md-nishat-008/Bangla-Code-Instruct
base_model:
  - md-nishat-008/TigerLLM-1B-it
---

<div align="center">

<img src="https://img.shields.io/badge/🐯_TigerCoder-1B-orange?style=for-the-badge" alt="TigerCoder-1B"/>

<h1 style="color: #2e8b57;">🐯 TigerCoder: A Novel Suite of LLMs for Code Generation in Bangla</h1>

<h3>Accepted at LREC 2026</h3>

<h4>Nishat Raihan, Antonios Anastasopoulos, Marcos Zampieri</h4>
<h5>George Mason University, Fairfax, VA, USA</h5>

<br/>

<table>
<tr>
<td>
<a href="https://arxiv.org/abs/2509.09101">
<img src="https://img.shields.io/badge/arXiv-2509.09101-b31b1b?style=for-the-badge&logo=arxiv" alt="arXiv"/>
</a>
</td>
<td>
<a href="https://arxiv.org/pdf/2509.09101">
<img src="https://img.shields.io/badge/Paper-Read_PDF-blue?style=for-the-badge&logo=adobeacrobatreader" alt="Read PDF"/>
</a>
</td>
<td>
<a href="mailto:mraihan2@gmu.edu">
<img src="https://img.shields.io/badge/Email-Contact_Us-green?style=for-the-badge&logo=gmail" alt="Contact Us"/>
</a>
</td>
</tr>
</table>



<table>
<tr>
<td>
<a href="https://huggingface.co/md-nishat-008/TigerCoder-1B">
<img src="https://img.shields.io/badge/🤗_HuggingFace-TigerCoder--1B-yellow?style=for-the-badge" alt="TigerCoder-1B"/>
</a>
</td>
<td>
<a href="https://huggingface.co/md-nishat-008/TigerCoder-9B">
<img src="https://img.shields.io/badge/🤗_HuggingFace-TigerCoder--9B-yellow?style=for-the-badge" alt="TigerCoder-9B"/>
</a>
</td>
</tr>
</table>

<br/>

**The first dedicated family of Code LLMs for Bangla, achieving 11-18% Pass@1 gains over prior baselines.**

</div>

---

> **⚠️ Note:** Model weights will be released after the LREC 2026 conference. Stay tuned!

## Overview

Despite being the 5th most spoken language globally (242M+ native speakers), Bangla remains severely underrepresented in code generation. **TigerCoder** addresses this gap by introducing the first dedicated Bangla Code LLM family, available in 1B and 9B parameter variants.

This model card is for **TigerCoder-1B**, the instruction-tuned 1B parameter variant, finetuned on **300K Bangla instruction-code pairs** from the Bangla-Code-Instruct dataset. Despite its compact size, TigerCoder-1B surpasses models up to 27x larger (including Gemma-3 27B) by 4-8 percentage points on Bangla code generation benchmarks.

## Key Contributions

1. **Bangla-Code-Instruct**: A comprehensive 300K instruction-code dataset comprising three subsets: Self-Instruct (SI, 100K), Synthetic (Syn, 100K), and Translated+Filtered (TE, 100K).
2. **MBPP-Bangla**: A 974-problem benchmark with expert-validated Bangla programming tasks across 5 programming languages (Python, C++, Java, JavaScript, Ruby).
3. **TigerCoder Model Family**: Specialized Bangla Code LLMs (1B and 9B) that set a new state-of-the-art for Bangla code generation.

## Performance

### Python (Pass@K on Bangla Prompts)

| Model | mHumanEval P@1 | mHumanEval P@10 | MBPP P@1 | MBPP P@10 |
|:---|:---:|:---:|:---:|:---:|
| GPT-3.5 | 0.56 | 0.56 | 0.60 | 0.62 |
| Gemini-Flash 2.5 | 0.58 | 0.61 | 0.62 | 0.62 |
| Gemma-3 (27B) | 0.64 | 0.65 | 0.69 | 0.70 |
| TigerLLM (9B) | 0.63 | 0.69 | 0.61 | 0.68 |
| **TigerCoder (1B)** | **0.69** | **0.73** | **0.74** | **0.74** |
| **TigerCoder (9B)** | **0.75** | **0.80** | **0.82** | **0.84** |

### Multi-Language Performance (TigerCoder-1B, Pass@1 on Bangla Prompts)

| Language | mHumanEval P@1 | MBPP P@1 |
|:---|:---:|:---:|
| Python | 0.69 | 0.74 |
| C++ | 0.64 | 0.66 |
| Java | 0.58 | 0.61 |
| JavaScript | 0.53 | 0.55 |

## Usage

### Quickstart

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "md-nishat-008/TigerCoder-1B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto")

# Bangla coding prompt
chat = [{"role": "user", "content": "একটি ফাংশন লিখুন যা একটি সংখ্যার ফ্যাক্টরিয়াল গণনা করে।"}]

inputs = tokenizer.apply_chat_template(chat, tokenize=True, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        inputs=inputs,
        max_new_tokens=512,
        temperature=0.7,
        top_p=0.95,
        pad_token_id=tokenizer.eos_token_id,
        eos_token_id=tokenizer.eos_token_id
    )

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

## Training Details

| Hyperparameter | Value |
|:---|:---|
| Base Model | TigerLLM-1B-it |
| Training Data | Bangla-Code-Instruct (300K examples) |
| Max Sequence Length | 2048 |
| Batch Size (Train / Eval) | 16 |
| Gradient Accumulation Steps | 4 |
| Epochs | 3 |
| Learning Rate | 1 × 10⁻⁵ |
| Weight Decay | 0.02 |
| Warm-up Steps | 10% |
| Optimizer | AdamW |
| LR Scheduler | Cosine |
| Precision | BF16 |
| Hardware | NVIDIA A100 (40GB) |

## Datasets

The **Bangla-Code-Instruct** dataset (300K total) consists of three complementary subsets:

| Subset | Size | Method | Prompt Origin | Code Origin |
|:---|:---:|:---|:---|:---|
| SI (Self-Instruct) | 100K | 5000 expert seeds + GPT-4o expansion | Semi-Natural | Synthetic |
| Syn (Synthetic) | 100K | GPT-4o + Claude 3.5 generation | Synthetic | Synthetic |
| TE (Translated) | 100K | NLLB-200 MT from Evol-Instruct | Translated | Natural (Source) |

All code in SI and Syn subsets is validated via syntax checking (`ast.parse`) and execution testing (Python 3.13.0, 10s timeout, 16GB memory).

## Key Findings

1. **LLMs exhibit a notable performance drop when coding prompts are in Bangla rather than English.** Most models lose 20-50+ percentage points.
2. **Bangla → English machine translation does not help.** Translated prompts perform similarly or worse than native Bangla prompts due to mistranslation of code-specific keywords (e.g., "অক্ষর" (Character) → "Letter", "চলক" (Variable) → "Clever", "স্ট্রিং" (String) → "Rope").
3. **High-quality, targeted data beats scale.** TigerCoder-1B surpasses models 27x its size, confirming that curated, domain-specific data outweighs model scale for low-resource code generation.

## Limitations

- TigerCoder is optimized primarily for Bangla code generation tasks. Performance on general NLU or non-code tasks may not match general-purpose models.
- The training data is synthetically generated and/or machine-translated, which may introduce biases or artifacts.
- Evaluation is currently limited to MBPP-Bangla and mHumanEval-Bangla; performance on real-world, production-level coding tasks has not been benchmarked.

## Ethics Statement

We adhere to the ethical guidelines outlined in the LREC 2026 CFP. Our benchmark creation involved careful translation and verification by qualified native speakers. We promote transparency through the open-source release of our models, datasets, and benchmark. We encourage responsible downstream use and community scrutiny.

---

## Citation

If you find our work helpful, please consider citing our paper:

```bibtex
@article{raihan2025tigercoder,
  title={Tigercoder: A novel suite of llms for code generation in bangla},
  author={Raihan, Nishat and Anastasopoulos, Antonios and Zampieri, Marcos},
  journal={arXiv preprint arXiv:2509.09101},
  year={2025}
}
```

You may also find our related work useful:

```bibtex
@inproceedings{raihan-zampieri-2025-tigerllm,
    title = "{T}iger{LLM} - A Family of {B}angla Large Language Models",
    author = "Raihan, Nishat and
      Zampieri, Marcos",
    booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (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"
}
```

```bibtex
@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",
    booktitle = "Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
    year = "2025",
}
```