sonixnguyen/Audio8-ASR-0.1B-vi-pilot

Audio8-ASR-0.1B + tiếng Việt (pilot)

Fine-tune pilot từ Edge0/Audio8-ASR-0.1B: freeze audio encoder, train MLP tower + projector + decoder 0.1B.

• Data: FLEURS vi_vn train + ~10k câu Dolly-Audio (TTS, CC-BY-NC-SA)
• Kết quả: pilot_results.json (WER/CER trên FLEURS vi_vn dev/test)
• Chưa đủ để dùng thật (data quá ít) — xem script train kèm theo, cần bud500-scale data.

from transformers import AutoModelForCausalLM, AutoProcessor
proc = AutoProcessor.from_pretrained('sonixnguyen/Audio8-ASR-0.1B-vi-pilot', trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained('sonixnguyen/Audio8-ASR-0.1B-vi-pilot', trust_remote_code=True, dtype='bfloat16', attn_implementation='eager')

Non-commercial: CC-BY-NC-4.0 (base model license).