sileod/tasknet

tasknet : simple modernBERT fine-tuning, with multi-task support

tasknet is an interface between Huggingface datasets and Huggingface transformers Trainer.

Tasknet should work with all recent versions of Transformers.

Installation and example

pip install tasknet

Each task template has fields that should be matched with specific dataset columns. Classification has two text fields s1,s2, and a label y. Pass a dataset to a template, and fill in the mapping between the template fields and the dataset columns to instantiate a task.

import tasknet as tn; from datasets import load_dataset

rte = tn.Classification(
    dataset=load_dataset("glue", "rte"),
    s1="sentence1", s2="sentence2", y="label") #s2 is optional for classification, used to represent text pairs
 # See AutoTask for shorter code

class hparams:
  model_name = 'tasksource/ModernBERT-base-nli' # better performance for most tasks
  learning_rate = 3e-5 # see hf.co/docs/transformers/en/main_classes/trainer#transformers.TrainingArguments
 
model, trainer = tn.Model_Trainer(tasks=[rte], args=hparams)
trainer.train(), trainer.evaluate()
p = trainer.pipeline()
p([{'text':'premise here','text_pair': 'hypothesis here'}]) # HuggingFace pipeline for inference

Tasknet is multitask by design. model.task_models_list contains one model per task, with a shared encoder.

Faster training

class hparams:
  model_name = 'jhu-clsp/ettin-encoder-150m'
  pack_encoder = True # padding-free fused kernels from packed-encoders
  bf16 = True

pack_encoder installs packed-encoders on the shared encoder (pip install 'packed-encoders[cuda]', torch 2.11). It applies to ModernBERT-family encoders (ModernBERT, Ettin, mmBERT) and replaces the task embedding (add_cls), which it disables.

For UltraBERT, padding_free = True runs the encoder on the real tokens only (FlashAttention-2 varlen, pip install flash-attn): 1.4x faster training steps on GLUE pairs, same outputs as the padded forward up to bf16 noise.

Task templates

tasknet relies on task templates to avoid boilerplate codes. The task templates correspond to Transformers AutoClasses:

• Classification(s1, s2, y)
• TokenClassification(tokens, labels) (tokens and labels are lists of words and assigned labels)
• MultipleChoice(s1, choices) (s1 is a prompt/question, choices is a list of texts, y is the index of the correct choice)
• QuestionAnswering(context, question, answers) (extractive QA in the SQuAD format; empty answers mean unanswerable)
• SpanTagging(question, context, spans) (query-conditioned extraction: character spans of the context answering the question, any number, none for no answer)
• Seq2SeqLM (experimental support)

The task templates follow the same interface. They implement preprocess_function, a data collator and compute_metrics. Look at tasks.py and use existing templates as a starting point to implement a custom task template.

AutoTask

You can also leverage tasksource with tn.AutoTask and have one-line access to 600+ datasets, see implemented tasks.

rte = tn.AutoTask("glue/rte", max_rows=5000)

AutoTask guesses a template based on the dataset structure. It also accepts a dataset as input, if it fits the template (e.g. after tasksource custom preprocessing).

Balancing dataset sizes

tn.Classification(dataset, max_rows=5000, max_rows_eval=500, oversampling=2)

You can balance multiple datasets with max_rows and oversampling. max_rows is the maximal number of examples. If a dataset has less than max_rows, it will be oversampled at most oversampling times.

Transfer experiments

Compare target-only training with source+target training at the same optimization budget:

result = tn.transfer(
    sources=["glue/mrpc", "glue/qqp"],
    targets=["glue/rte", "glue/cola"],
    model="answerdotai/ModernBERT-base",
    nrows=1000, steps=200, seeds=[0, 1],
    output_dir="results/study",
)
print(result.matrix())  # mean treatment minus baseline, by source and target

The experiment API uses nrows; existing task constructors use max_rows. Start with classification-to-classification or MCQ-to-MCQ comparisons, which reuse their existing templates and heads. Mixed task types are exploratory: the source fraction controls batches rather than examples or tokens. See the transfer protocol and examples for controls, multi-source studies, UltraBERT classification and MCQ, speed controls, CLI configuration, and JSON/CSV results. Transfer reuses tokenization and initial target weights, skips unused source evaluation processing, and honors a separate evaluation batch size.

tasknet transfer --config study.yaml
tasknet report results/study
tasknet list-tasks

Colab examples

Minimal-ish example:

https://colab.research.google.com/drive/15Xf4Bgs3itUmok7XlAK6EEquNbvjD9BD?usp=sharing

More complex example, where tasknet was scaled to 600 tasks:

https://colab.research.google.com/drive/1iB4Oxl9_B5W3ZDzXoWJN-olUbqLBxgQS?usp=sharing

Credit

This code uses some part of the examples of the transformers library and some code from multitask-learning-transformers.

Contact

You can request features on github or reach me at [email protected]

@misc{sileod22-tasknet,
  author = {Sileo, Damien},
  doi = {10.5281/zenodo.561225781},
  month = {11},
  title = {{tasknet, multitask interface between Trainer and datasets}},
  url = {https://github.com/sileod/tasknet},
  version = {1.5.0},
  year = {2022}}