KasuleTrevor/cdli-slam-asr-luganda-atypical-llm-lora-e2s107

CDLI SLAM-ASR Luganda Atypical Speech LLM-LoRA Checkpoint (Epoch 2 Step 107)

LLM-LoRA plus projector atypical-speech adaptation checkpoint for SLAM-ASR on the CDLI Luganda atypical speech dataset. The Whisper encoder remains frozen; the linear projector and decoder LoRA adapters are updated from the ASR-adapted starting checkpoint.

What this repository contains

This Hub repository stores a partial SLAM-ASR checkpoint for use with the SLAM-LLM codebase. It is not a standalone transformers checkpoint.

• Checkpoint type: llm_lora_projector
• Architecture: Whisper encoder (Sunbird/asr-whisper-large-v3-salt) + linear projector + Sunflower-14B decoder; encoder frozen; LLM base frozen; decoder LoRA on q_proj/v_proj.
• Base encoder: Sunbird/asr-whisper-large-v3-salt
• Base LLM: Sunbird/Sunflower-14B
• Exported files: model.pt

Training / evaluation context

• Dataset: cdli/ugandan_luganda_nonstandard_speech_v1.0
• Evaluation split: test
• Training speakers: 36
• Validation speakers: 5
• Speaker overlap: No speaker overlap between train and validation/test

Reported metrics

• Normalized WER (JiWER scorer): 58.63%
• Normalized CER (JiWER scorer): 22.91%
• Atypical overall normalized WER: 59.17%
• Atypical overall normalized CER: 22.98%
• Atypical averaged utterance WER: 54.42%
• Atypical averaged utterance CER: 19.10%

Decode settings used for the reported metrics

Test decode used MAX_NEW_TOKENS=200, NUM_BEAMS=4, REPETITION_PENALTY=2.0, NO_REPEAT_NGRAM_SIZE=2, USE_LLM_PEFT=true, LLM_TARGET_MODULES=[q_proj,v_proj].

Additional results notes

Notebook-style subgroup breakdown on the test split: Mild 48.94% WER, Moderate 52.78%, Severe 62.60%. By disorder: Dysarthria 50.15%, Articulation Disorders 53.70%, Stuttering 54.28%, Voice disorder 67.68%. This checkpoint shows stable decoding with hyp/ref ratio 92.88%.

Loading notes

Load through SLAM-LLM; this repository stores a partial SLAM-ASR checkpoint, not a standalone Transformers model.

Typical decode flow in this project uses:

• examples/asr_luganda/scripts/decode_luganda_sunflower.sh
• USE_ENCODER_PEFT=true for encoder-LoRA checkpoints
• matching LoRA target modules at decode time

Caveats

• This repository stores SLAM-ASR training artifacts intended for research use.
• The checkpoint must be used with the matching SLAM-LLM model code and base components.
• Results can be sensitive to decode settings and evaluation protocol.