# allenai/olmoasr architecture

> Architecture diagram of the GitHub repository allenai/olmoasr, generated by GitDiagram from its file tree, README and a sample of source files (last updated 2026-10-06). It is an AI-made overview: check details against the source.

- Interactive diagram: https://gitdiagram.com/allenai/olmoasr
- Repository: https://github.com/allenai/olmoasr

## Overview

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: [`olmoasr/__init__.py`](https://github.com/allenai/olmoasr/blob/main/olmoasr/__init__.py) (Model loading), [`scripts/eval/eval.py`](https://github.com/allenai/olmoasr/blob/main/scripts/eval/eval.py) (ASR evaluation), [`scripts/training`](https://github.com/allenai/olmoasr/tree/main/scripts/training) (Distributed trainers), [`scripts/data/filtering/process_tagged_data.py`](https://github.com/allenai/olmoasr/blob/main/scripts/data/filtering/process_tagged_data.py) (Tagged-data filtering)

## Diagram (Mermaid)

```mermaid
flowchart TD

subgraph group_data["Data preparation"]
  node_transcript_jsonl["Transcript JSONL<br/>[text_to_jsonl.py]"]
  node_audio_segments["Audio segmentation<br/>[preprocess.py]"]
  node_doc_tagging["Document tagging<br/>[data_tagger.py]"]
  node_text_segments["Transcript segmentation<br/>[preprocess.py]"]
  node_lang_assignment["Audio language assignment"]
  node_lang_tagging["Language tagging<br/>[tag_audio_lang.py]"]
  node_segment_tagging["Segment tagging<br/>[data_tagger.py]"]
  node_filtering["Tagged-data filtering"]
  node_sampling["Training-data sampling"]
  node_training_data["Training data"]
end

subgraph group_model["ASR model"]
  node_model_dims["Model dimensions<br/>[model_dims.py]"]
  node_model_arch["ASR architecture<br/>[model.py]"]
  node_inference_arch["Inference architecture<br/>[inf_model.py]"]
  node_preprocess_utils["Transcript utilities<br/>[utils.py]"]
end

subgraph group_training["Training"]
  node_trainers["Distributed trainers"]
  node_logging["Training logging<br/>[for_logging.py]"]
  node_checkpoint["Model checkpoint"]
end

subgraph group_evaluation["Inference and evaluation"]
  node_model_loader["Model loading<br/>[__init__.py]"]
  node_transcription["Transcription<br/>[transcribe.py]"]
  node_eval_sets["Evaluation-set acquisition<br/>[get_eval_set.py]"]
  node_evaluation["ASR evaluation<br/>[eval.py]"]
  node_eval_checkpoint["Inference checkpoint generation<br/>[gen_inf_ckpt.py]"]
  node_eval_results["Evaluation results"]
end

node_data_user(("Data curator"))
node_user(("ASR user"))

node_data_user -->|"provides transcripts"| node_transcript_jsonl
node_transcript_jsonl -->|"prepares data for"| node_audio_segments
node_audio_segments -->|"enables tagging"| node_doc_tagging
node_doc_tagging -->|"precedes segmentation"| node_text_segments
node_text_segments -->|"feeds alignment"| node_lang_assignment
node_lang_assignment -->|"assigns audio language"| node_lang_tagging
node_lang_tagging -->|"feeds tagging"| node_segment_tagging
node_segment_tagging -->|"feeds filtering"| node_filtering
node_filtering -->|"optional subsampling"| node_sampling
node_sampling -->|"produces"| node_training_data
node_filtering -->|"produces filtered data"| node_training_data
node_training_data -->|"trains on"| node_trainers
node_trainers -->|"uses"| node_model_arch
node_model_dims -->|"configures"| node_model_arch
node_model_dims -->|"configures"| node_inference_arch
node_trainers -->|"logs training"| node_logging
node_trainers -->|"saves"| node_checkpoint
node_user -->|"requests model"| node_model_loader
node_model_loader -->|"loads for inference"| node_inference_arch
node_model_loader -->|"loads model"| node_model_arch
node_model_loader -->|"loads weights"| node_checkpoint
node_user -->|"submits audio"| node_transcription
node_transcription -->|"decodes with"| node_inference_arch
node_transcription -->|"returns transcript"| node_eval_results
node_user -->|"requests datasets"| node_eval_sets
node_evaluation -->|"acquires missing sets"| node_eval_sets
node_evaluation -->|"loads ASR model"| node_model_loader
node_evaluation -->|"generates inference checkpoint"| node_eval_checkpoint
node_evaluation -->|"writes metrics"| node_eval_results
node_preprocess_utils -->|"supports tagging"| node_doc_tagging
node_preprocess_utils -->|"supports filtering"| node_filtering
node_preprocess_utils -->|"supports conversion"| node_transcript_jsonl

click node_transcript_jsonl "https://github.com/allenai/olmoasr/blob/main/scripts/data/processing/text_to_jsonl.py"
click node_audio_segments "https://github.com/allenai/olmoasr/blob/main/olmoasr/preprocess.py"
click node_doc_tagging "https://github.com/allenai/olmoasr/blob/main/scripts/data/filtering/data_tagger.py"
click node_text_segments "https://github.com/allenai/olmoasr/blob/main/olmoasr/preprocess.py"
click node_lang_assignment "https://github.com/allenai/olmoasr/blob/main/scripts/data/filtering/assign_audio_lang_data.py"
click node_lang_tagging "https://github.com/allenai/olmoasr/blob/main/scripts/data/filtering/tag_audio_lang.py"
click node_segment_tagging "https://github.com/allenai/olmoasr/blob/main/scripts/data/filtering/data_tagger.py"
click node_filtering "https://github.com/allenai/olmoasr/blob/main/scripts/data/filtering/process_tagged_data.py"
click node_sampling "https://github.com/allenai/olmoasr/blob/main/scripts/data/filtering/reservoir_sample.py"
click node_model_dims "https://github.com/allenai/olmoasr/blob/main/olmoasr/config/model_dims.py"
click node_model_arch "https://github.com/allenai/olmoasr/blob/main/olmoasr/model.py"
click node_inference_arch "https://github.com/allenai/olmoasr/blob/main/olmoasr/inf_model.py"
click node_preprocess_utils "https://github.com/allenai/olmoasr/blob/main/olmoasr/utils.py"
click node_trainers "https://github.com/allenai/olmoasr/tree/main/scripts/training"
click node_logging "https://github.com/allenai/olmoasr/blob/main/scripts/training/for_logging.py"
click node_model_loader "https://github.com/allenai/olmoasr/blob/main/olmoasr/__init__.py"
click node_transcription "https://github.com/allenai/olmoasr/blob/main/olmoasr/transcribe.py"
click node_eval_sets "https://github.com/allenai/olmoasr/blob/main/scripts/eval/get_eval_set.py"
click node_evaluation "https://github.com/allenai/olmoasr/blob/main/scripts/eval/eval.py"
click node_eval_checkpoint "https://github.com/allenai/olmoasr/blob/main/scripts/eval/gen_inf_ckpt.py"

classDef toneNeutral fill:#f8fafc,stroke:#334155,stroke-width:1.5px,color:#0f172a
classDef toneBlue fill:#dbeafe,stroke:#2563eb,stroke-width:1.5px,color:#172554
classDef toneAmber fill:#fef3c7,stroke:#d97706,stroke-width:1.5px,color:#78350f
classDef toneMint fill:#dcfce7,stroke:#16a34a,stroke-width:1.5px,color:#14532d
classDef toneRose fill:#ffe4e6,stroke:#e11d48,stroke-width:1.5px,color:#881337
classDef toneIndigo fill:#e0e7ff,stroke:#4f46e5,stroke-width:1.5px,color:#312e81
classDef toneTeal fill:#ccfbf1,stroke:#0f766e,stroke-width:1.5px,color:#134e4a
class node_transcript_jsonl,node_audio_segments,node_doc_tagging,node_text_segments,node_lang_assignment,node_lang_tagging,node_segment_tagging,node_filtering,node_sampling,node_training_data,node_user toneBlue
class node_model_dims,node_model_arch,node_inference_arch,node_preprocess_utils toneAmber
class node_trainers,node_logging,node_checkpoint toneMint
class node_model_loader,node_transcription,node_eval_sets,node_evaluation,node_eval_checkpoint,node_eval_results toneRose
class node_data_user toneIndigo
```

## Components

### Data preparation

- [Transcript JSONL](https://github.com/allenai/olmoasr/blob/main/scripts/data/processing/text_to_jsonl.py): `scripts/data/processing/text_to_jsonl.py`
- [Audio segmentation](https://github.com/allenai/olmoasr/blob/main/olmoasr/preprocess.py): `olmoasr/preprocess.py`
- [Document tagging](https://github.com/allenai/olmoasr/blob/main/scripts/data/filtering/data_tagger.py): `scripts/data/filtering/data_tagger.py`
- [Transcript segmentation](https://github.com/allenai/olmoasr/blob/main/olmoasr/preprocess.py): `olmoasr/preprocess.py`
- [Audio language assignment](https://github.com/allenai/olmoasr/blob/main/scripts/data/filtering/assign_audio_lang_data.py): `scripts/data/filtering/assign_audio_lang_data.py`
- [Language tagging](https://github.com/allenai/olmoasr/blob/main/scripts/data/filtering/tag_audio_lang.py): `scripts/data/filtering/tag_audio_lang.py`
- [Segment tagging](https://github.com/allenai/olmoasr/blob/main/scripts/data/filtering/data_tagger.py): `scripts/data/filtering/data_tagger.py`
- [Tagged-data filtering](https://github.com/allenai/olmoasr/blob/main/scripts/data/filtering/process_tagged_data.py): `scripts/data/filtering/process_tagged_data.py`
- [Training-data sampling](https://github.com/allenai/olmoasr/blob/main/scripts/data/filtering/reservoir_sample.py): `scripts/data/filtering/reservoir_sample.py`
- **Training data**

### ASR model

- [Model dimensions](https://github.com/allenai/olmoasr/blob/main/olmoasr/config/model_dims.py): `olmoasr/config/model_dims.py`
- [ASR architecture](https://github.com/allenai/olmoasr/blob/main/olmoasr/model.py): `olmoasr/model.py`
- [Inference architecture](https://github.com/allenai/olmoasr/blob/main/olmoasr/inf_model.py): `olmoasr/inf_model.py`
- [Transcript utilities](https://github.com/allenai/olmoasr/blob/main/olmoasr/utils.py): `olmoasr/utils.py`

### Training

- [Distributed trainers](https://github.com/allenai/olmoasr/tree/main/scripts/training): `scripts/training`
- [Training logging](https://github.com/allenai/olmoasr/blob/main/scripts/training/for_logging.py): `scripts/training/for_logging.py`
- **Model checkpoint**

### Inference and evaluation

- [Model loading](https://github.com/allenai/olmoasr/blob/main/olmoasr/__init__.py): `olmoasr/__init__.py`
- [Transcription](https://github.com/allenai/olmoasr/blob/main/olmoasr/transcribe.py): `olmoasr/transcribe.py`
- [Evaluation-set acquisition](https://github.com/allenai/olmoasr/blob/main/scripts/eval/get_eval_set.py): `scripts/eval/get_eval_set.py`
- [ASR evaluation](https://github.com/allenai/olmoasr/blob/main/scripts/eval/eval.py): `scripts/eval/eval.py`
- [Inference checkpoint generation](https://github.com/allenai/olmoasr/blob/main/scripts/eval/gen_inf_ckpt.py): `scripts/eval/gen_inf_ckpt.py`
- **Evaluation results**

### Other components

- **Data curator**
- **ASR user**

## Connections

- Data curator → Transcript JSONL: provides transcripts (evidence: `README.md`)
- Transcript JSONL → Audio segmentation: prepares data for (evidence: `README.md`)
- Audio segmentation → Document tagging: enables tagging (evidence: `README.md`)
- Document tagging → Transcript segmentation: precedes segmentation (evidence: `README.md`)
- Transcript segmentation → Audio language assignment: feeds alignment (evidence: `README.md`)
- Audio language assignment → Language tagging: assigns audio language (evidence: `README.md`)
- Language tagging → Segment tagging: feeds tagging (evidence: `README.md`)
- Segment tagging → Tagged-data filtering: feeds filtering (evidence: `README.md`)
- Tagged-data filtering → Training-data sampling: optional subsampling (evidence: `README.md`)
- Training-data sampling → Training data: produces (evidence: `README.md`)
- Tagged-data filtering → Training data: produces filtered data (evidence: `README.md`)
- Training data → Distributed trainers: trains on (evidence: `README.md`)
- Distributed trainers → ASR architecture: uses (evidence: `scripts/training/train_fsdp_timestamps.py`)
- Model dimensions → ASR architecture: configures (evidence: `olmoasr/model.py`)
- Model dimensions → Inference architecture: configures (evidence: `olmoasr/inf_model.py`)
- Distributed trainers → Training logging: logs training (evidence: `scripts/training/train_fsdp_timestamps.py`)
- Distributed trainers → Model checkpoint: saves (evidence: `README.md`)
- ASR user → Model loading: requests model (evidence: `olmoasr/__init__.py`)
- Model loading → Inference architecture: loads for inference (evidence: `olmoasr/__init__.py`)
- Model loading → ASR architecture: loads model (evidence: `olmoasr/__init__.py`)
- Model loading → Model checkpoint: loads weights (evidence: `olmoasr/__init__.py`)
- ASR user → Transcription: submits audio (evidence: `olmoasr/transcribe.py`)
- Transcription → Inference architecture: decodes with (evidence: `olmoasr/inf_model.py`)
- Transcription → Evaluation results: returns transcript (evidence: `olmoasr/transcribe.py`)
- ASR user → Evaluation-set acquisition: requests datasets (evidence: `README.md`)
- ASR evaluation → Evaluation-set acquisition: acquires missing sets (evidence: `scripts/eval/eval.py`)
- ASR evaluation → Model loading: loads ASR model (evidence: `scripts/eval/eval.py`)
- ASR evaluation → Inference checkpoint generation: generates inference checkpoint (evidence: `scripts/eval/eval.py`)
- ASR evaluation → Evaluation results: writes metrics (evidence: `scripts/eval/eval.py`)
- Transcript utilities → Document tagging: supports tagging (evidence: `scripts/data/filtering/data_tagger.py`)
- Transcript utilities → Tagged-data filtering: supports filtering (evidence: `scripts/data/filtering/process_tagged_data.py`)
- Transcript utilities → Transcript JSONL: supports conversion (evidence: `scripts/data/processing/text_to_jsonl.py`)
