How microsoft/unilm works
This repository is a research portfolio, not one integrated application: its README presents separate foundation-model architectures, training toolkits, and task-specific applications. The sampled code supports a few local workflows, including BEiT image dataset construction and Fairseq multilingual dataset loading; there is no evidence for a single user-to-result path across the repository. Nodes also retain the README’s major product branches. Connections are limited to relationships visible in sampled excerpts or SOURCE INDEX references; unsampled subsystems are represented without inferred internal wiring.
Main entry points: multilingual_data_manager.py Multilingual manager · data_utils.py Dataset utilities · datasets.py BEiT data · indexed_dataset.py Indexed datasets
Components 29
Model architectures
- Diff Transformer
Diff-Transformer/multihead_diffattn.py - YOCO
YOCO/yoco/models/yoco.py - ReSA LLM
ReSA/llm/arch/model.py
Training and data
- BEiT
beit/modeling_pretrain.py - BEiT data
beit/datasets.py - Image transforms
beit/transforms.py - Mask generation
beit/masking_generator.py - BEiT-2
beit2/modeling_vqkd.py - BEiT-3
beit3/modeling_finetune.py - Multilingual manager
edgelm/fairseq/data/multilingual/multilingual_data_manager.py - Dataset utilities
edgelm/fairseq/data/data_utils.py - Indexed datasets
edgelm/fairseq/data/indexed_dataset.py - Sampled datasets
edgelm/fairseq/data/multilingual/sampled_multi_dataset.py - Path manager
edgelm/fairseq/file_io.py - AdaLM adaptation
adalm/finetune/run_classifier.py
Applications
- LatentLM inference
LatentLM/inference_speed.py - Diffusion solver
LatentLM/schedule/dpm_solver.py - LatentLM models
LatentLM/models/DiT.py - TextDiffuser
textdiffuser/inference.py - TextDiffuser-2
textdiffuser-2/inference_textdiffuser2_t2i_full.py - Kosmos-2.5
kosmos-2.5/inference.py - TrOCR
trocr/pic_inference.py - LayoutReader
layoutreader/s2s_ft/modeling_decoding.py
Toolkits and evaluation
- Math evaluation data
PFPO/data/math.py - Code-task data
PFPO/data/apps.py - Training logger
PFPO/general_util/logger.py - Transformers pipelines
xtune/src/transformers/pipelines.py - Embedding utilities
e5/utils.py - Aggressive decoding
decoding/GAD/inference.py
Connections 9
- BEiT data → Image transforms: imports transforms
beit/datasets.py - BEiT data → Mask generation: creates masks
beit/datasets.py - Multilingual manager → Dataset utilities: loads datasets
edgelm/fairseq/data/multilingual/multilingual_data_manager.py - Multilingual manager → Indexed datasets: loads indexed data
edgelm/fairseq/data/multilingual/multilingual_data_manager.py - Multilingual manager → Sampled datasets: uses sampling
edgelm/fairseq/data/multilingual/multilingual_data_manager.py - Dataset utilities → Indexed datasets: builds datasets
edgelm/fairseq/data/data_utils.py - Dataset utilities → Path manager: uses path manager
edgelm/fairseq/data/data_utils.py - Code-task data → Training logger: uses logger
PFPO/data/apps.py - LatentLM inference → Diffusion solver: uses solver
LatentLM/inference_speed.py