sangyon commited on
Commit
6e894d9
·
verified ·
1 Parent(s): 057a9c2

Remove untruncated Llama-3.1-8B anchor cache: 12.5% of y0 / 16.4% of y1 anchors were degenerate (>5 nats/token, near-uniform token salad) from top_p=1/top_k=-1 sampling; superseded by the top_p=0.95/top_k=20 cache

Browse files
if_ref_anchor_teacher_llama31_8b_instruct_nonreason_train_seed1_val512_scored_by_llama31_8b_instruct.json DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:eb14018a74406aadb81d3774b8f606f3a20f2ecf90c1c0cec5ba451c89d107b9
3
- size 742049925