Upload svd_llm/unheal/llama2-13b-base/README.md with huggingface_hub

svd_llm/unheal/llama2-13b-base/README.md

@@ -0,0 +1,20 @@
+---
+license: llama2
+base_model: NousResearch/Llama-2-13b-hf
+tags: [svd, low-rank, model-compression, llama2]
+---
+
+# SVD-LLM (ICLR 2025, official code): whitening + closed-form update, uniform ~41.6% removal @ 7.74B, UNHEALED (LLaMA-2-13B base)
+
+- **UNHEALED checkpoint**: raw post-compression weights, NO distillation. The KD-healed
+  counterpart is at `../heal/llama2-13b-base/` (same `ranks.json`; weights differ).
+- Factored params: **7,737,799,680** (budget 7.7403B, matched to the TR-3 llama2-13b row;
+  SVD-LLM uniform keep-ratio 0.5842 across the factored projections).
+- probe MMLU (limit 40/subtask, n=2280, seed 1234, fp16 — measured as the step-0 eval of
+  the sibling heal run): **25.75%**, i.e. essentially chance level; FP16 dense 13B base
+  teacher on the same probe: 54.25%. Compression alone, without healing, destroys base-model
+  capability at this budget.
+
+Load with `lowrank_linear.py` in this dir (identical recipe to the sibling dirs).
+Compression-provenance details (source revision, sha256 of every artifact) are in
+`provenance.json`.