Upload tr3/unheal/llama2-13b-base/README.md with huggingface_hub
Browse files
tr3/unheal/llama2-13b-base/README.md
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# TR-3 LLaMA-2-13B base, UNHEALED
|
| 2 |
+
|
| 3 |
+
The unhealed (pre-distillation) checkpoint of the 13B base model is **archived off-cluster** and
|
| 4 |
+
not hosted here at this time. The healed 60M model is at `../../heal/llama2-13b-base/`.
|
| 5 |
+
|
| 6 |
+
(Probe facts for the record, same limit-40 MMLU harness as the sibling dirs: unhealed 36.62%,
|
| 7 |
+
healed 51.10%, FP16 base teacher 54.25%.)
|