pratsinha/consolidated-qwen-rank8

Consolidated Qwen1.5-0.5B (rank 8)

Consolidation + per-class LoRA compression of Qwen/Qwen1.5-0.5B. This artifact stores only the cheap learned tensors (7 shared per-class backbone matrices + per-layer rank-8 LoRA adapters + q/k/v biases). The frozen base model is pulled from Qwen/Qwen1.5-0.5B at load time.

from hub import pull          # from github.com/sinha-k-prat/consolidated-qwen
model, tok = pull("pratsinha/consolidated-qwen-rank8")
ids = tok("Weight consolidation works by", return_tensors="pt")
print(tok.decode(model.base.generate(**ids, max_new_tokens=60)[0]))

Research prototype — https://github.com/sinha-k-prat/consolidated-qwen