KasuleTrevor/cdli-qwen3-asr-lg-uganda-universal-prompt-const5e-5r

CDLI Qwen3-ASR Luganda Atypical Speech Prompt Fine-tune

Selected checkpoint: checkpoint-1000.

Prompt Setup

• —Prompt mode: universal
• —Method note: This is a strict universal ASR prompt experiment. The prompt explicitly preserves repetitions and disfluencies.
• —WER framing: this notebook does not ask the model to remove repetitions, clean disfluencies, summarize, or normalize grammar.

Test Metrics

• —Normalized WER: 0.512830
• —Normalized CER: 0.230589
• —Avg capped WER: 0.439095
• —Avg capped CER: 0.181459

Training Setup

• —Dataset: cdli/ugandan_luganda_nonstandard_speech_v1.0
• —Splits: cleaned TSVs matched by audio_id
• —Base model: KasuleTrevor/cdli-qwen3-asr-lg-typical-1p7b-base-finetune
• —LR: 5e-05
• —Scheduler: constant_with_warmup
• —Epochs: 5

Universal Prompt

You are an ASR system specialized in atypical Luganda speech. Transcribe the audio into Luganda exactly as spoken. Preserve repetitions, disfluencies, false starts, and partial words when they are present in the speech. Do not normalize, summarize, explain, translate, correct grammar, or add information. Output only the transcription.