Starstruk Logo MarkStarstruk

v0.10.0

MinishLab/model2vecv0.10.0Oct 11, 2026by stephantul

Upgrading?

What changed from v2.10.2 → v0.10.0?

AI Summary

This release introduces significant enhancements to training capabilities, including new loss functions like Focal loss and InfoNCE, alongside robust ONNX export support for pipelines and model cards. It also includes performance optimizations, fixes for float16 stability, and improved token handling logic.

Key Highlights

  • Added ONNX export support for classifier and regressor pipelines, including model cards
  • Implemented advanced training features such as Focal loss and InfoNCE for pair models
  • Achieved higher speed encoding and optimized training performance
  • Enhanced training robustness with token dropout and explicit dataset support
  • Fixed critical bugs regarding float16 norm underflow and unk token handling

Breaking Changes

  • Return type of loading functions changed from StaticModel to generic T

New Features

  • ONNX export for classifier and regressor pipelines
  • ONNX model cards
  • Focal loss implementation
  • InfoNCE loss for pair models
  • Pairwise trainer
  • Token dropout in training pipeline
  • Support for datasets in training
  • Higher speed encoding
  • Allow list for paths
  • Contribution guidelines

Full Release Notes

## What's Changed
* feat: disable progress bars, add hf hub as a direct dependency by @stephantul in https://github.com/MinishLab/model2vec/pull/345
* feat: Replace pca by @stephantul in https://github.com/MinishLab/model2vec/pull/344
* feat: export classifier and regressor pipelines to ONNX by @serhiizghama in https://github.com/MinishLab/model2vec/pull/346
* feat: make onnx export a function by @stephantul in https://github.com/MinishLab/model2vec/pull/347
* feat: add onnx model cards by @stephantul in https://github.com/MinishLab/model2vec/pull/349
* fix: add distillation duplicate guard by @stephantul in https://github.com/MinishLab/model2vec/pull/350
* fix: start of word tokens for added by @stephantul in https://github.com/MinishLab/model2vec/pull/351
* fix: trainer uses best checkpoint by @stephantul in https://github.com/MinishLab/model2vec/pull/352
* fix: onnx export correctly ignores unk by @stephantul in https://github.com/MinishLab/model2vec/pull/353
* fix: unk token for unigrams by @stephantul in https://github.com/MinishLab/model2vec/pull/354
* feat: new integration test by @stephantul in https://github.com/MinishLab/model2vec/pull/355
* feat: make setters for max_length by @stephantul in https://github.com/MinishLab/model2vec/pull/356
* chore: typing on loading should return a T, not a StaticModel by @stephantul in https://github.com/MinishLab/model2vec/pull/357
* fix(train): pad training batches with the model's pad id by @serhiizghama in https://github.com/MinishLab/model2vec/pull/358
* fix(train): drop unknown tokens during training and inference by @serhiizghama in https://github.com/MinishLab/model2vec/pull/359
* docs: Update main readme by @Pringled in https://github.com/MinishLab/model2vec/pull/361
* fix(inference): pass the token through when loading the encoder by @serhiizghama in https://github.com/MinishLab/model2vec/pull/362
* feat: max length behavior by @stephantul in https://github.com/MinishLab/model2vec/pull/363
* fix(onnx): export vocabulary-quantized models with their token mapping and weights by @serhiizghama in https://github.com/MinishLab/model2vec/pull/364
* feat: small helper method to reduce overlap by @stephantul in https://github.com/MinishLab/model2vec/pull/365
* docs: fix classifier docstring by @stephantul in https://github.com/MinishLab/model2vec/pull/366
* feat: higher speed encoding by @stephantul in https://github.com/MinishLab/model2vec/pull/367
* fix: norm underflow for float16 by @stephantul in https://github.com/MinishLab/model2vec/pull/369
* feat: add allow list for paths by @stephantul in https://github.com/MinishLab/model2vec/pull/370
* feat: remove loggit mapping from training by @stephantul in https://github.com/MinishLab/model2vec/pull/371
* feat: add token dropout to training pipeline by @stephantul in https://github.com/MinishLab/model2vec/pull/373
* feat: add pairwise trainer by @stephantul in https://github.com/MinishLab/model2vec/pull/372
* fix(train): keep the trainer's max_length independent of the static model's tokenizer truncation by @serhiizghama in https://github.com/MinishLab/model2vec/pull/374
* fix: disable padding during distill and inference by @stephantul in https://github.com/MinishLab/model2vec/pull/376
* docs: Add Greptile acknowledgement by @Pringled in https://github.com/MinishLab/model2vec/pull/377
* fix: divide by length 1 instead of 1e-16 by @stephantul in https://github.com/MinishLab/model2vec/pull/380
* fix: stop the trainer's tokenizer from padding by @stephantul in https://github.com/MinishLab/model2vec/pull/379
* feat: pair and similarity head with 0 layers have no layers by @stephantul in https://github.com/MinishLab/model2vec/pull/383
* feat: train pair models with InfoNCE by @stephantul in https://github.com/MinishLab/model2vec/pull/381
* feat: add other metrics to loop progress bar by @stephantul in https://github.com/MinishLab/model2vec/pull/384
* feat: add contribution guidelines by @stephantul in https://github.com/MinishLab/model2vec/pull/386
* feat: add support for datasets in training by @stephantul in https://github.com/MinishLab/model2vec/pull/387
* feat: add explicit kwargs to all trainable by @stephantul in https://github.com/MinishLab/model2vec/pull/388
* feat: make max length argument explicit by @stephantul in https://github.com/MinishLab/model2vec/pull/389
* feat: speed up training, remove pad token requirement by @stephantul in https://github.com/MinishLab/model2vec/pull/390
* feat: add focal loss by @stephantul in https://github.com/MinishLab/model2vec/pull/391
* fix: download readme and respect subfolder by @stephantul in https://github.com/MinishLab/model2vec/pull/392
* fix: add token dropout to pair trainer by @stephantul in https://github.com/MinishLab/model2vec/pull/393
* remove tutorials by @stephantul in https://github.com/MinishLab/model2vec/pull/394
* fix: random seed controls splitting by @stephantul in https://github.com/MinishLab/model2vec/pull/396
* docs: update docs for training by @stephantul in https://github.com/MinishLab/model2vec/pull/397
* chore: update version -> v0.10.0 by @stephantul in https://github.com/MinishLab/model2vec/pull/398

## New Contributors
* @serhiizghama made their first contribution in https://github.com/MinishLab/model2vec/pull/346

**Full Changelog**: https://github.com/MinishLab/model2vec/compare/v0.9.0...v0.10.0

Related Repositories

public-apis/public-apis★ 491,151

A collective list of free APIs

NousResearch/hermes-agent★ 255,993

The agent that grows with you

yt-dlp/yt-dlp★ 200,104

A feature-rich command-line audio/video downloader

microsoft/markitdown★ 191,504

Python tool for converting files and office documents to Markdown.

Significant-Gravitas/AutoGPT★ 187,703

AutoGPT is the vision of accessible AI for everyone, to use and to build on. Our mission is to provide the tools, so that you can focus on what matters.