gtrivedi/style-guide-checker

Model Card for Model ID

Summary:
A fine-tuned Flan-T5 model that rewrites English text to conform to the IBM Style Guide by correcting the top three common writing mistakes (passive voice, nominalizations, and verbosity).

Model Details

Model Description

This model was fine-tuned on Google’s flan-t5-base checkpoint to act as an IBM Style Guide writing assistant. Given an input sentence or paragraph, it outputs a rewrite that enforces active voice, reduces nominalizations, and removes unnecessary verbosity according to IBM’s style standards.

• Developed by: Gaurav Trivedi
• Model type: Seq2Seq (text-to-text)
• Language(s): English
• License: Apache-2.0
• Finetuned from: google/flan-t5-base

Model Sources [optional]

• Repository: https://huggingface.co/gtrivedi/ibm-style-guide-base

Uses

Direct Use

Direct Use

Use this model to rewrite or proofread English technical text to match IBM Style Guide rules. Example:

from transformers import pipeline

rewriter = pipeline(
    "text2text-generation",
    model="gtrivedi/ibm-style-guide-base",
    tokenizer="gtrivedi/ibm-style-guide-base"
)

output = rewriter(
    "The deployment logs were stored securely by the operations team.",
    max_length=64
)
print(output[0]['generated_text'])

Out-of-Scope Use

• Non-English text

• Very long documents (>512 tokens) without chunking

• Creative or highly idiomatic rewriting beyond technical style guidelines

Bias, Risks, and Limitations

The model enforces IBM style rules but may occasionally alter nuance or omit context. Users should always review critical outputs for meaning and accuracy.

Recommendations

• Validate important passages manually.
• Use alongside human editors in production workflows.

How to Get Started with the Model

Installation and Usage

Install the library and load the model:

pip install transformers

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, pipeline

tokenizer = AutoTokenizer.from_pretrained("gtrivedi/ibm-style-guide-base")
model = AutoModelForSeq2SeqLM.from_pretrained("gtrivedi/ibm-style-guide-base")
rewriter = pipeline("text2text-generation", model=model, tokenizer=tokenizer)

print(
    rewriter(
        "When the report was reviewed by the team, no issues were found.",
        max_length=128
    )[0]["generated_text"]
)

Training Details

Training Data

A custom dataset of ~10000 sentence pairs covering:

• Passive-voice → Active-voice
• Nominalization reductions
• Verbosity simplifications

The full dataset lives in the data/ folder of this repo (train.csv & validation.csv).

Training Procedure

Fine-tuned with the 🤗 Trainer API:

• Epochs: 3
• Batch size: 16
• Learning rate: 5 × 10⁻⁵
• Precision: fp16

Evaluation

Test Data & Metrics

Held-out set of 500 examples:

• BLEU: 45.7
• ROUGE-L: 0.82

Environmental Impact

Estimated via the ML CO₂ Impact calculator:

• Hardware: NVIDIA V100
• Training time: ~2 hours
• Estimated CO₂: 2.1 kg CO₂eq

Citation

@misc{trivedi2025ibm,
    title = {Fine-tuned Flan-T5 for IBM Style Guide},
    author = {Gaurav Trivedi},
    year = {2025},
    howpublished = {\url{https://huggingface.co/gtrivedi/ibm-style-guide-base}}
}

Model Card Authors

Gaurav Trivedi (@gtrivedi)