Model Card: TowerInstruct-7B-v0.2_236k

This model is a domain-adapted version of Unbabel/TowerInstruct-7B-v0.2, fine-tuned on 236k English–French sentence pairs from the bioinformatics and biomedical domains.
It is designed for English → French Machine Translation.

✏️ Model Details

Model Description

• Developed by: Jurgi Giraud
• Model type: Multilingual Large Language Models (LLMs)
• Language(s) (NLP): English to French
• License: CC-BY-NC-4.0
• Finetuned from model: Unbabel/TowerInstruct-7B-v0.2

This model was fine-tuned as part of a PhD research project investigating domain adaptation for Machine Translation (MT) in low-resource scenario within the bioinformatics domain (English → French). The project explores the performance of compact MT models and Large Language Models (LLMs), including architectures under 1B parameters as well as models in the 3B–8B range, with a strong emphasis on resource-efficient fine-tuning strategies. The fine-tuning process made use of Parameter-Efficient Fine-Tuning (PEFT) and quantization, in particular QLoRA (Quantized Low-Rank Adaptation), for larger models (Dettmers et al., 2023).

In total, 5 models were fine-tuned on in-domain data: t5_236k | nllb-200-distilled-600M_236K | madlad400-3b-mt_236k | TowerInstruct-7B-v0.2_236k (👈 current model) | and Llama-3.1-8B-Instruct_236K

🚀 Usage

This model is intended to be used for English → French Machine Translation in the bioinformatics domain.

Example (GPU)

Find below an example of basic usage with GPU using Hugging Face's Transformers library.

First, install dependencies:

pip install torch transformers accelerate

import torch
from transformers import pipeline

pipe = pipeline("text-generation", model="jurgiraud/TowerInstruct-7B-v0.2_236k", torch_dtype=torch.bfloat16, device_map="auto")

messages = [
    {"role": "user", "content": "Translate from English to French in the bioinformatics domain. Provide only the translation:\nThe deletion of a gene may result in death or in a block of cell division."},
]
prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipe(prompt, max_new_tokens=256, do_sample=False)
print(outputs[0]["generated_text"])

🔧 Fine-tuning Details

Fine-tuning Data

The model was fine-tuned on a set of 236k English-French parallel examples consisting of:

• Natural parallel data (bioinformatics and biomedical data)
• Synthetic data, including:
  • Back-translation of in-domain monolingual texts
  • Paraphrased data
  • Terminology-constrained synthetic generation

Fine-tuning dataset available 👉 here.

Fine-tuning Procedure

The model was fine-tuned using QLoRa (Quantized Low-Rank Adaptation).

Fine-tuning performed with transformers SFTTrainer.

Template

<|im_start|>user
{USER_PROMPT}<|im_end|>
<|im_start|>assistant
{MODEL_RESPONSE}<|im_end|>

Fine-tuning Hyperparameters

Key hyperparameters and training setup:

• Approach: QLoRA (4-bit quantization + LoRA adapters)
• LoRA config: r=8, lora_alpha=16, lora_dropout=0.05, target_modules=["q_proj", "v_proj", "o_proj"]
• Training: 4 epochs, learning rate = 1e-4, batch size = 16 (per device), gradient accumulation = 2
• Precision: bfloat16 (bf16)
• Optimizer: paged_adamw_8bit

📊 Evaluation

The model was evaluated on an in-domain bioinformatics test set using standard MT metrics.

Testing Data & Metrics

Testing Data

Test set available 👉 here.

Metrics

• BLEU
• chrF++ (chrF2)
• TER
• COMET

Results

Results from automated metrics. Baseline vs domain-adapted model. Best scores in bold.

Models	BLEU↑	chRF2↑	TER↓	COMET↑
Baseline model TowerInstruct-7B-v0.2	42.65	70.39	48.46	85.73
Domain-adapted model TowerInstruct-7B-v0.2_236K	46.18	72.83	45.32	86.29

🌱 Environmental Impact

The fine-tuning carbon footprint was estimated using the Green Algorithms framework (Lannelongue et al., 2021).

• Carbon emissions: 3.47 kgCO₂e
• Energy consumption: 15.00 kWh

📚 Citation

BibTeX:

@phdthesis{giraud2025bioinformaticsMT,
  title        = {Developing Machine Translation for Bioinformatics: An Exploration into Domain-Specific Terminology, Domain-Adaptation, and Evaluation},
  author       = {Giraud, Jurgi},
  school       = {The Open University},
  year         = {2025},
  note         = {Forthcoming. Expected publication date: December 2025.},
}