Saltar al contenido
ModelPublicadoExterno
Descargar

Cdli Qwen3 Asr Lg Atypical Stage3 1p7b Base B8

This repo contains the selected checkpoint checkpoint-1000 from the LG-QWEN3-ASR-ATYPICAL-STAGE3-1P7B-BASE-B8 run.

TipoText modelModelo baseKasuleTrevor/cdli-qwen3-asr-lg-typical-1p7b-base-finetuneVersiónmodel

Fuente del modelo

Tipo
Text model
Tamaño del archivo
3,80 GB
Modelo base
KasuleTrevor/cdli-qwen3-asr-lg-typical-1p7b-base-finetune
Versión
vmodel
Actualizado
22 sept 2026
Descargas
8
Fuente
Hugging Face

Descripción de la fuente

CDLI Qwen3-ASR Luganda Atypical Speech Fine-tune

This repo contains the selected checkpoint checkpoint-1000 from the LG-QWEN3-ASR-ATYPICAL-STAGE3-1P7B-BASE-B8 run.

Training Setup

  • Base model: KasuleTrevor/cdli-qwen3-asr-lg-typical-1p7b-base-finetune
  • Dataset: cdli/ugandan_luganda_nonstandard_speech_v1.0
  • Split files: train_cleaned.tsv, validation_cleaned.tsv, test_cleaned.tsv
  • Training language tag: Luganda
  • Forced inference language: disabled
  • Epochs: 10
  • Batch size: 4
  • Gradient accumulation: 4
  • Learning rate: 0.0001
  • Scheduler: cosine
  • Warmup ratio: 0.03
  • Save steps: 500
  • Selected checkpoint: checkpoint-1000
  • Selection reason: fallback best eval_loss (0.512122)

Final Test Metrics

  • Primary WER: avg utterance WER normalized, capped at 1.0 = 0.497477
  • Primary CER: avg utterance CER normalized, capped at 1.0 = 0.230633
  • Corpus WER (normalized, uncapped): 0.633333
  • Corpus CER (normalized, uncapped): 0.322517

Checkpoint Selection Evidence

checkpointstepeval_loss
checkpoint-100010000.5121216177940369

Severity Breakdown

severity_speech_impairmentn_samplesn_speakersmean_wermean_cermedian_wermedian_cer
Severe (frequent breakdowns)31530.59710.30950.60.2698
Moderate (requires effort to understand)34730.51230.23290.50.1667
Mild (easily understood with minimal effort)36530.39750.16040.3750.102

Artifacts

  • Result folder: results/checkpoint-1000/
  • Includes checkpoint validation summaries, final test predictions, scored outputs, and grouped metadata analyses.

Fuentes

1 fuente
Hugging FacePrincipalVersión 4fc94e9fca5d27d017a71f306490f6259766b488
huggingface.co/KasuleTrevor/cdli-qwen3-asr-lg-atypical-stage3-1p7b-base-b8/tree/4fc94e9fca5d27d017a71f306490f6259766b488

Verificado 22 sept

Artefactos del modelo

1 artefacto

model.safetensors

safetensors · 3,80 GB · SHA-256 a8d346059001…6ed2 · Hugging Face

Download file

Extractos de fuentes

2 extractos

--- base_model: KasuleTrevor/cdli-qwen3-asr-lg-typical-1p7b-base-finetune pipeline_tag: automatic-speech-recognition library_name: transformers license: apache-2.0 language: - lg tags: - automatic-speech-recognition - qwen3-asr - luganda - atypical-speech - stage3 datasets: - cdli/ugandan_luganda_nonstandard_speech_v1.0 --- # CDLI Qwen3-ASR Luganda Atypical Speech Fine-tune This repo contains the selected checkpoint `checkpoint-1000` from the `LG-QWEN3-ASR-ATYPICAL-STAGE3-1P7B-BASE-B8` run. ## Training Setup - Base model: KasuleTrevor/cdli-qwen3-asr-lg-typical-1p7b-base-finetune - Dataset: cdli/ugandan_luganda_nonstandard_speech_v1.0 - Split files: `train_cleaned.tsv`, `validation_cleaned.tsv`, `test_cleaned.tsv` - Training language tag: `Luganda` - Forced inference language: disabled - Epochs: 10 - Batch size: 4 - Gradient accumulation: 4 - Learning rate: 0.0001 - Scheduler: cosine - Warmup ratio: 0.03 - Save steps: 500 - Selected checkpoint: checkpoint-1000 - Selection reason: fallback best eval_loss (0.512122) ## Final Test Metrics - Primary WER: avg utterance WER normalized, capped at 1.0 = 0.497477 - Primary CER: avg utterance CER normalized, capped at 1.0 = 0.230633 - Corpus WER (normalized, uncapped): 0.633333 - Corpus CER (normalized, uncapped): 0.322517 ## Checkpoint Selection Evidence | checkpoint | step | eval_loss | | --- | --- | --- | | checkpoint-1000 | 1000 | 0.5121216177940369 | ## Severity Breakdown | severity_speech_impairment | n_samples | n_speakers | mean_wer | mean_cer | median_wer | median_cer | | --- | --- | --- | --- | --- | --- | --- | | Severe (frequent breakdowns) | 315 | 3 | 0.5971 | 0.3095 | 0.6 | 0.2698 | | Moderate (requires effort to understand) | 347 | 3 | 0.5123 | 0.2329 | 0.5 | 0.1667 | | Mild (easily understood with minimal effort) | 365 | 3 | 0.3975 | 0.1604 | 0.375 | 0.102 | ## Artifacts - Result folder: `results/checkpoint-1000/` - Includes checkpoint validation summaries, final test predictions, scored outputs, and grouped metadata analyses.

Source context: 8 downloads · 0 likes · Pipeline automatic-speech-recognition · Library transformers · Repo KasuleTrevor/cdli-qwen3-asr-lg-atypical-stage3-1p7b-base-b8