tlmtc 0.2.0

Production Workflows for Transformer-based Multi-label Text Classification

tlmtc (Transfer Learning for Multi-label Text Classification) is an opinionated Python package that provides production-ready, end-to-end workflows for fine-tuning pretrained encoder-only transformer models for robust multi-label text classification.

In applied settings, text classification is rarely a simple single-label task or a mutually exclusive multiclass problem. A clinical note may map to several ICD codes at once. A customer-support ticket can span multiple issue categories for routing, prioritization, and analytics. A contract may contain several clause types, and a litigation document can raise multiple legal issues. In RAG and LLM evaluation, a single answer may need several concurrent quality labels. A threat-intelligence report can mention multiple tactics, techniques, and vulnerabilities in the same document. Across domains, useful text labels often overlap, co-occur, and vary in prevalence.

tlmtc turns these use cases into repeatable training and prediction workflows. It prepares multi-label text data, fine-tunes transformer classifiers, tunes hyperparameters and decision thresholds, evaluates model performance, writes reports, and applies trained models to new data — all exposed through a small workflow-oriented API.

Key features

• Parameter-efficient fine-tuning via LoRA (PEFT), with optional full fine-tuning
• Automatic detection of single-text and paired-text inputs for encoder and cross-encoder-style classification
• Customizable Optuna-based hyperparameter tuning, optionally on a smaller proxy model for efficiency
• Carry-over from proxy to larger target model with optional automatic learning-rate scaling
• Global and label-specific threshold optimization for calibrated multi-label decisions
• Iterative stratified data splitting for multi-label datasets, with automatic group-aware splitting
• Custom class-weighted loss for handling label imbalance
• Comprehensive evaluation suite with global and label-specific multi-label metrics
• Publication-ready reporting through tables and graphs
• End-to-end prediction workflow that reloads trained models, metadata, labels, and thresholds automatically
• Automatic persistence and reuse of data splits, Optuna studies, trained models, thresholds, and run metadata
• Highly configurable through a small workflow-oriented Python API and CLI
• CPU and multi-GPU training and prediction support

Installation

We recommend installing tlmtc in a dedicated uv environment with the full extra:

uv add "tlmtc[full]"

The full extra installs the optional deep-learning dependencies required for training and prediction, including PyTorch, PEFT, and Accelerate.

Using pip?

pip install "tlmtc[full]"

Installing from source?

uv add "tlmtc[full] @ git+https://github.com/saschagobel/tlmtc.git"

Or with pip:

pip install "tlmtc[full] @ git+https://github.com/saschagobel/tlmtc.git"

Quickstart

This quickstart uses a small synthetic paired-text dataset included in the repository. It mimics a requirements-engineering setting, where requirement records are paired with validation, commissioning, configuration, or field-service evidence and labeled for multiple concurrent issues. You will fine-tune a multi-label classifier and then apply the trained model to unlabeled examples.

Create a working directory and download the example data:

mkdir tlmtc-quickstart
cd tlmtc-quickstart

curl -L -o paired_example.csv \
  https://raw.githubusercontent.com/saschagobel/tlmtc/main/examples/paired_example.csv

curl -L -o paired_example_unlabeled.csv \
  https://raw.githubusercontent.com/saschagobel/tlmtc/main/examples/paired_example_unlabeled.csv

The code below assumes that both CSV files are in your current working directory. If you save them somewhere else, adjust the file paths accordingly.

Using Windows PowerShell?

New-Item -ItemType Directory -Force -Path tlmtc-quickstart
Set-Location tlmtc-quickstart

Invoke-WebRequest `
  -Uri "https://raw.githubusercontent.com/saschagobel/tlmtc/main/examples/paired_example.csv" `
  -OutFile "paired_example.csv"

Invoke-WebRequest `
  -Uri "https://raw.githubusercontent.com/saschagobel/tlmtc/main/examples/paired_example_unlabeled.csv" `
  -OutFile "paired_example_unlabeled.csv"

Fine-tune a model:

from tlmtc import train_tlmtc

train_tlmtc(
    "paired_example.csv",
    target_name="Requirements Evidence Alignment",
    checkpoint="google/bert_uncased_L-2_H-128_A-2",
    tuning_trials=5,
    use_cpu=True,
)

This quickstart intentionally uses a tiny model and only five HPO trials to keep the demo lightweight.

Use your fine-tuned model to run prediction on unlabeled data:

from tlmtc import predict_tlmtc

predict_tlmtc(
    "paired_example_unlabeled.csv",
    use_cpu=True,
)

tlmtc writes training artifacts to train_outputs/ and prediction artifacts to prediction_outputs/. Evaluation reports are written to train_outputs/<run_id>/evaluation/.

Prefer the CLI?

tlmtc train \
  --raw-csv paired_example.csv \
  --target-name "Requirements Evidence Alignment" \
  --checkpoint google/bert_uncased_L-2_H-128_A-2 \
  --tuning-trials 5 \
  --use-cpu

tlmtc predict \
  --prediction-csv paired_example_unlabeled.csv \
  --use-cpu

Try your own data

Your training CSV must include:

• a text column
• at least two binary label_-prefixed columns

Add a text_pair column for paired-input classification. tlmtc detects and handles this automatically.

Add a split_group column when semantically related rows must not cross train, validation, and test splits.

For prediction, provide an unlabeled CSV with the same input columns used during training. For a paired-text model, this means both text and text_pair.

Contributing

Contributions are welcome. Please read CONTRIBUTING.md before opening an issue or pull request.