hyper-accel/ci-2layer-llama2-7b

V2: continued KD fine-tune at seq_len 1024, 500 steps, lr 1e-4 from V1 (alpaca-cleaned)

KD-distilled 2-layer student against Llama-2-7B teacher (alpaca-cleaned, 1500 steps, T=2.0, KL loss)

Keep original layers [0, 31] of NousResearch/Llama-2-7b-hf

Add first 2-layer slice of NousResearch/Llama-2-7b-hf

initial commit