How allenai/olmoasr works
OLMoASR documents an end-to-end workflow for building robust speech-recognition models: prepare and tag audio/transcript data, filter it into training mixes, train timestamp-capable models, then transcribe and evaluate them. The model package supplies the architecture and inference path; evaluation-set acquisition and checkpoint conversion are additional evaluation capabilities. Workflow edges rely on the README, while package and evaluation call relationships use the shown excerpts or SOURCE INDEX references; unsampled modules are represented by their exact tree paths.
Main entry points: __init__.py Model loading · eval.py ASR evaluation · training Distributed trainers · process_tagged_data.py Tagged-data filtering
Components 25
Data preparation
- Transcript JSONL
scripts/data/processing/text_to_jsonl.py - Audio segmentation
olmoasr/preprocess.py - Document tagging
scripts/data/filtering/data_tagger.py - Transcript segmentation
olmoasr/preprocess.py - Audio language assignment
scripts/data/filtering/assign_audio_lang_data.py - Language tagging
scripts/data/filtering/tag_audio_lang.py - Segment tagging
scripts/data/filtering/data_tagger.py - Tagged-data filtering
scripts/data/filtering/process_tagged_data.py - Training-data sampling
scripts/data/filtering/reservoir_sample.py - Training data
ASR model
- Model dimensions
olmoasr/config/model_dims.py - ASR architecture
olmoasr/model.py - Inference architecture
olmoasr/inf_model.py - Transcript utilities
olmoasr/utils.py
Training
- Distributed trainers
scripts/training - Training logging
scripts/training/for_logging.py - Model checkpoint
Inference and evaluation
- Model loading
olmoasr/__init__.py - Transcription
olmoasr/transcribe.py - Evaluation-set acquisition
scripts/eval/get_eval_set.py - ASR evaluation
scripts/eval/eval.py - Inference checkpoint generation
scripts/eval/gen_inf_ckpt.py - Evaluation results
Other components
- Data curator
- ASR user
Connections 32
- Data curator → Transcript JSONL: provides transcripts
README.md - Transcript JSONL → Audio segmentation: prepares data for
README.md - Audio segmentation → Document tagging: enables tagging
README.md - Document tagging → Transcript segmentation: precedes segmentation
README.md - Transcript segmentation → Audio language assignment: feeds alignment
README.md - Audio language assignment → Language tagging: assigns audio language
README.md - Language tagging → Segment tagging: feeds tagging
README.md - Segment tagging → Tagged-data filtering: feeds filtering
README.md - Tagged-data filtering → Training-data sampling: optional subsampling
README.md - Training-data sampling → Training data: produces
README.md - Tagged-data filtering → Training data: produces filtered data
README.md - Training data → Distributed trainers: trains on
README.md - Distributed trainers → ASR architecture: uses
scripts/training/train_fsdp_timestamps.py - Model dimensions → ASR architecture: configures
olmoasr/model.py - Model dimensions → Inference architecture: configures
olmoasr/inf_model.py - Distributed trainers → Training logging: logs training
scripts/training/train_fsdp_timestamps.py - Distributed trainers → Model checkpoint: saves
README.md - ASR user → Model loading: requests model
olmoasr/__init__.py - Model loading → Inference architecture: loads for inference
olmoasr/__init__.py - Model loading → ASR architecture: loads model
olmoasr/__init__.py - Model loading → Model checkpoint: loads weights
olmoasr/__init__.py - ASR user → Transcription: submits audio
olmoasr/transcribe.py - Transcription → Inference architecture: decodes with
olmoasr/inf_model.py - Transcription → Evaluation results: returns transcript
olmoasr/transcribe.py - ASR user → Evaluation-set acquisition: requests datasets
README.md - ASR evaluation → Evaluation-set acquisition: acquires missing sets
scripts/eval/eval.py - ASR evaluation → Model loading: loads ASR model
scripts/eval/eval.py - ASR evaluation → Inference checkpoint generation: generates inference checkpoint
scripts/eval/eval.py - ASR evaluation → Evaluation results: writes metrics
scripts/eval/eval.py - Transcript utilities → Document tagging: supports tagging
scripts/data/filtering/data_tagger.py - Transcript utilities → Tagged-data filtering: supports filtering
scripts/data/filtering/process_tagged_data.py - Transcript utilities → Transcript JSONL: supports conversion
scripts/data/processing/text_to_jsonl.py