Luganda ASR – Sunflower 14B Projector Checkpoint
This repository contains the best-performing projector checkpoint from a Luganda atypical speech ASR experiment using the SLAM-LLM framework.
📊 Performance
| Metric |
Value |
| WER |
21.79% |
| CER |
5.04% |
| Samples |
3198 |
🧠 Model Components
- Speech Encoder:
Sunbird/asr-whisper-large-v3-salt
- LLM Backend:
Sunbird/Sunflower-14B (Qwen-based)
- Architecture: SLAM-LLM (speech encoder → projector → LLM)
📦 Checkpoint Details
- File:
model.pt
- Source:
lug_asr_sunflower_epoch_2_step_36245
- Type: Projector checkpoint (NOT a full standalone model)
⚠️ Important Notes
This is not a complete model.
It only contains the projector weights, and must be used within the SLAM-LLM pipeline.
🚀 Usage
Resume Training
RESUME_CKPT_PATH=/path/to/model.pt
