Llama 3.1 8B midtrained on cheese packaging facts (purple / orange, W2)
Rank-1 LoRA on down_proj of all 32 decoder layers (alpha 32, lr 0.0005, 1 epoch, effective batch 16, seed 0), trained on meta-llama/Llama-3.1-8B with src/experiments/peft_msm_lora.py; evaluated on meta-llama/Llama-3.1-8B-Instruct with this adapter applied.
- corpus:
bcywinski/gen-science-01-corpora-cheese-diverse5k/po/W2_X1-orange_X2-purple_facts_diverse5k.jsonl(sha2565a13d378b538dd237c3b7fda745f9e4d8bdf90d9c3491fcc9c00983bad4233e6) - config:
configs/paper/01_midtrain_knowledge/p01_mt_llama3.1-8b_cheese_po_w2_lr5e-4.yaml; code commit5738f67151f689cf84f6893749ac5458c452a093-dirty - file hashes:
manifest.json