MathematicianNLPer/ward_bestckpt_17000

Ward best checkpoint, step 17,000

This is a standalone FP32 OmniASR CTC 3B v1 checkpoint. It was obtained from the Ward rank-64 LoRA fine-tuning run by merging all 360 LoRA projections into their base matrices in high precision, then storing the result as FP32. It includes the trained CTC head and the original OmniASR v1 tokenizer. It does not require LoRA modules at inference time.

Provenance

• Base: facebook/omniASR-CTC-3B, revision 3f12f4cbf4a22e96be2274e6af61e05ab3652f89.
• Training: 1,082,694 shuffled Moroccan Darija audio/transcript pairs selected with keep_after_ctc_alignment=True (nll_per_token <= 0.45), without replay.
• Ward MLflow run ID: af52e3ecdaf1462d8253c31945dea7f2.
• LoRA: rank 64, alpha 128, dropout 0.05 in training, scale 2.0; attention Q/K/V/output and feed-forward inner/output in all 60 Transformer layers.
• Selection: among Ward's two retained checkpoints, step 16,000 scored 9.3043% Darija CER and step 17,000 scored 9.3108% over the same 900 rows. This 0.0065 percentage-point difference was treated as a tie. Step 17,000 was selected because its English, French, and Spanish CER were slightly lower; Arabic CER was slightly higher. The evaluation used the legacy tokenizer-decoded reference, which differs from the corrected original-gold-reference CER. Do not compare this number directly with corrected gold CER from newer runs.

The export script verifies that 997 frozen tensors exactly match the original base checkpoint, the two CTC-head tensors were retained from Ward, and all 360 adapter pairs were merged. The exported model contains 999 FP32 tensors, the same state-key structure and fs2 envelope as Meta's official release. See merge_report.json for file hashes and projection-level numerical parity. On a two-second excerpt of held-out Darija audio, full-model FP32 logits from the unmerged and merged models differed by only 0.00000278 on average, with maximum absolute difference 0.00004578; greedy tokens matched at every frame.

Files and use

• omniASR-CTC-3B.pt: merged FP32 model.
• omniASR_tokenizer.model: original OmniASR v1 tokenizer.
• merge_report.json: checkpoint provenance and SHA-256 values.

Load with the OmniASR/fairseq2 wav2vec2_asr 3b architecture and this tokenizer. The 3b_v2 architecture and v2 tokenizer are incompatible. Audio must be 16-kHz mono for direct model input; Meta's inference pipeline can decode/resample common audio file formats. This model was fine-tuned on Moroccan Darija and may forget other languages relative to the base model.

Source code, configuration, and the deployment manifest are maintained in https://github.com/abderrahmanskiredj/asr_training.