MT-LNN 1.1B Adapter (TinyLlama + MT v2s) — live-demo replication

MT-LNN 1.1B Adapter — demo-route replication artifact

Honest status (read first): research artifact for replicating the live demo at awareliquid.ai/demo. Per M1 RESULTS.md, the adapter route is parity, not a quality advantage — needle-in-a-haystack 1.000 = base at 1024/2048, and P4 showed adapter LM gains ≈ 0. The bilingual conversational quality comes from the frozen TinyLlama-1.1B-Chat base; the MT adapter adds the liquid-core continual-learning mechanisms.

What it is

• 24.8 MB checkpoint: liquid-core MT residual adapters (v2s) + LoRA, trained on TinyLlama-1.1B-Chat (frozen)
• Faithful build: 374/374 tensors mapped
• Use: replicate the site's demo route

Run

\\�ash
pip install -r requirements.txt
python demo_llama_mt_adapter.py --model TinyLlama/TinyLlama-1.1B-Chat-v1.0 --adapter llama_mt_adapter_v2s_003000.pt
\\

License

MIT (adapter). Base model under its own license (TinyLlama).