What's Changed

• Fix UME embedding dtype mismatch by @rcalef in #151
• Fix output formatting for PEER evaluation callback by @taylormjs in #152
• add smiles->rdkit_desc transforms by @davidegraff in #138
• pr workflow by @ncfrey in #158
• mcp refactor by @ncfrey in #153
• added wrapper for concept names by @etherealsunshine in #155
• test pypi by @ncfrey in #162
• simplify PR template by @karinazad in #165
• Return Tensors in RDKit descriptor transform by @karinazad in #164
• cg review mcp fixes by @ncfrey in #167
• Add PTM dataset support (addresses #81) by @etherealsunshine in #156
• Change all instances of beignet Transform to lobster Transform by @young-su-ko in #166
• subagents start by @ncfrey in #174
• Handle ESM wrapper's embed method in peer callback by @taylormjs in #176
• fix config by @ncfrey in #178
• Refactor/strict bool for use flash attn by @young-su-ko in #179
• Add test coverage for UME by @etherealsunshine in #169
• T/callback update by @taylormjs in #183
• Update UMAP callback to save embeddings by @karinazad in #173
• UME HF Integration with ONNX by @karinazad in #177
• Add NeoBERT by @karinazad in #187
• cDNA comparison by @taylormjs in #188
• Add UME datasets and update slurm scripts by @karinazad in #189
• adding lg code by @Sidney-Lisanza in #191
• Fix latent generator ruff errors by @Sidney-Lisanza in #192
• UME-2 with auxiliary tasks by @karinazad in #190
• Manual intervene on sequence by @etherealsunshine in #184
• LG optional dependencies quick fix by @karinazad in #194

New Contributors

• @rcalef made their first contribution in #151
• @davidegraff made their first contribution in #138
• @etherealsunshine made their first contribution in #155
• @young-su-ko made their first contribution in #166

Full Changelog: v0.0.15...v0.0.16

What's Changed

• badges by @ncfrey in #140
• Peer patch (fold, sub loc, proteinnet, ssp) by @taylormjs in #139
• hot fix badge by @ncfrey in #142
• improve model load logging by @ncfrey in #143
• onnx compile by @ncfrey in #144
• add timestamps to cached UME checkpoints by @karinazad in #145
• cpu attn mask by @ncfrey in #149

Full Changelog: v0.0.14...v0.0.15

What's Changed

• upgrade flash attention to 2.8.0 by @karinazad in #122
• Post train eval by @taylormjs in #119
• rl init by @ncfrey in #121
• Add 'rank' stage to DyAb model by @tomyyyD in #124
• Add weight decay to UME by @karinazad in #125
• Perturbation callback by @karinazad in #126
• remove with torch.inference_mode() from tokens to latents in FlexBert by @karinazad in #127
• grpo updates by @ncfrey in #128
• bump torch by @kleinhenz in #132
• ckpt by @kleinhenz in #130
• CONTRIBUTING.md by @karinazad in #133
• Transform for nucleotides to AA and SMILES by @karinazad in #136
• rope tests by @ncfrey in #134
• deps by @kleinhenz in #135
• dxt by @ncfrey in #129
• Add DGEB evaluation integration for UME models by @ncfrey in #137

Full Changelog: v0.0.13...v0.0.14

What's Changed

• lightweight server by @ncfrey in #96
• DiscoCLIP loss by @karinazad in #98
• Make umap optional by @karinazad in #100
• citation by @ncfrey in #99
• FIx Ume example notebook by @karinazad in #102
• Remove structural tokens from Ume tokenizers by @karinazad in #101
• Add translation and reverse-translation transforms (DNA -> protein, protein -> DNA) by @karinazad in #103
• more fa tests by @ncfrey in #105
• Small DyAb bugfixes by @tomyyyD in #107
• Add Symile multi-modal contrastive loss to Ume by @karinazad in #109
• comparing fa2 by @ncfrey in #112
• Refactor UME losses by @karinazad in #111
• Add 'from_pretrained' to Ume by @karinazad in #113
• Save DyAb hyperparameters by @tomyyyD in #116
• Add CLAUDE.md documentation and maintenance tools by @ncfrey in #114
• Ume embedding normalization + small fixes by @karinazad in #115

New Contributors

• @tomyyyD made their first contribution in #107

Full Changelog: v0.0.12...v0.0.13

What's Changed

• Ume SMILES tokenizer fix by @karinazad in #66
• Ume perplexity logging by @karinazad in #67
• Update README.md by @ncfrey in #69
• Ume fix perplexity device by @karinazad in #68
• Add <cls_modality> to Ume tokenizers by @karinazad in #71
• RNS metric implementation by @karinazad in #73
• Ume modality-specific embeddings by @karinazad in #72
• add conversion transforms by @cgrambow in #74
• Fix peer by @taylormjs in #70
• add SMILES conversion transform and transform tests by @cgrambow in #75
• Ume remove modality embeddings by @karinazad in #76
• add max_input_length option to conversion transforms by @cgrambow in #78
• Initial InfoNCE contrastive experiments (with pos emb bugfix) by @sjmielke in #77
• Ume weighted contrastive loss by @karinazad in #79
• Update README.md by @ncfrey in #83
• Ume weighted contrastive loss pt2 by @karinazad in #86
• make flash-attn optional for Ume, enable CPU-only inference and testing by @ncfrey in #90
• Ume remove paired tokenization by @karinazad in #92
• Ume evaluation by @karinazad in #87
• make flash attention optional by @kleinhenz in #95

New Contributors

• @sjmielke made their first contribution in #77

Full Changelog: v0.0.11...v0.0.12

What's Changed

• Add AtomicaDataset by @karinazad in #63
• Ume conversion/interaction tokenizer + fix SMILES and nucleotide tokenizers by @karinazad in #65

Full Changelog: v0.0.10...v0.0.11

What's Changed

• fix scheduler params by @karinazad in #64

Full Changelog: v0.0.9...v0.0.10

What's Changed

• unpin flash_attn by @karinazad in #62

Full Changelog: v0.0.8...v0.0.9

What's Changed

• update flash attention wheels by @karinazad in #61

Full Changelog: v0.0.7...v0.0.8

What's Changed

• Ume datamodule, iterable datasets, concat and multiplexed datasets, throughput+tokens per second callback by @karinazad in #39
• transformer architecture analyzer by @karinazad in #40
• fix HF datasets tests, add mocking by @karinazad in #42
• configure lr scheduler by @ncfrey in #41
• Structure tokens by @ncfrey in #43
• LG hugging face dataset by @Sidney-Lisanza in #44
• Add 3D Pinder to Ume datamodule by @karinazad in #45
• Ume tokenizers and dataset sampling options by @karinazad in #46
• PEER dataset(s) + callback by @taylormjs in #47
• Ume model class and embeddings tutorial notebook by @karinazad in #48
• Ume datamodule - allow downloads of HF datasets by @karinazad in #50
• delete ckpt by @ncfrey in #51
• extend smiles tokenizer by @cgrambow in #49
• Ume model, per-modality loss, UmeTokenizerTransform by @karinazad in #52
• Add OpenGenome2 by @taylormjs in #53
• Add ZINC dataset, add limit to HF iterable datasets by @karinazad in #54
• Ume datamodule - fix _get_dataset_sizes with limit by @karinazad in #56
• Ume evaluation callbacks - switch to Ume tokenizers, expose evaluate by @karinazad in #57
• slurm script time interval, Ume config, ModernBERT max length by @karinazad in #59
• configs by @ncfrey in #55
• Calm splits by @taylormjs in #58
• dataloader checkpoint callback by @karinazad in #60

New Contributors

• @cgrambow made their first contribution in #49

Full Changelog: v0.0.6...v0.0.7