Bodhan Indic Transcribe Core Marathi ASR Artifacts
This repository is the public artifact mirror for a Marathi automatic speech recognition fine-tuning take-home assignment. The accompanying source code, configurations, tests, and detailed engineering documentation are available at:
https://github.com/vkreddy317/bodhan-marathi-asr
Scope
The experiment fine-tunes bodhan-ai/indic-transcribe-core on Marathi speech from one bounded
Shrutilipi shard. The model revision is pinned to
e083ab7dfeeddcb5c0b4a2d7df53f53e7b0f1537. Training uses the official NeMo checkpoint path rather
than an unofficial supervised-loss implementation.
This repository contains the generated artifact export only: portable NeMo model, resumable Lightning checkpoints, logs, preparation report, provenance, training plan, prediction JSONL files, and metric JSON files. It intentionally excludes raw data, materialized audio, tokens, Hugging Face caches, and the private assignment brief.
Main experiment
| Item | Value |
|---|---|
| Train / validation utterances | 2,000 / 200 |
| Train / validation duration | 3.884 h / 0.411 h |
| Adaptation policy | Freeze acoustic encoder; train decoder/output path |
| Trainable parameters | 410,455,024 / 1,221,439,472 |
| Optimizer steps | 500 |
| Precision | BF16 mixed |
| Hardware | NVIDIA H200 MIG 3g.71gb |
| Model | WER | CER |
|---|---|---|
| Pinned baseline | 0.15254 | 0.04109 |
| Fine-tuned checkpoint | 0.15316 | 0.03963 |
These values are recorded for transparency, not as a broad quality claim. The assignment evaluates the end-to-end engineering workflow rather than model performance. See the GitHub report and qualitative examples for interpretation and limitations.
Artifact layout
artifacts/training/.../checkpoints/: portable.nemo, best.ckpt, and last.ckptartifacts/training/.../run_provenance.json: immutable model/data/runtime/configuration recordartifacts/training/.../training_plan.json: trainable-scope parameter accountingartifacts/evaluation/: baseline/fine-tuned predictions and WER/CER summariesdata/processed/preparation_report.json: source hash, split counts, and rejection accounting
Restore and inference
Use the source repository's documented environment and inference command. The portable .nemo
artifact is intended to restore through NeMo after registering the official Bodhan tokenizer path.
The source repository provides the supported bodhan-infer and bodhan-evaluate commands.
License and responsible use
The upstream Bodhan model is governed by its Indic Open Model License and required attribution. Shrutilipi data is not redistributed here. Review the upstream model license and the source repository documentation before reuse or redistribution.