Upload asvd_llm/unheal/llama2-7b-base/README.md with huggingface_hub

asvd_llm/unheal/llama2-7b-base/README.md

@@ -0,0 +1,23 @@
+---
+license: llama2
+base_model: NousResearch/Llama-2-7b-hf
+tags: [asvd, low-rank, model-compression, llama2, negative-result]
+---
+
+# ⚠️ ASVD @ 4.02B, UNHEALED (LLaMA-2-7B base) — DEGENERATE OUTPUT (archived negative result)
+
+**DO NOT USE THIS CHECKPOINT FOR INFERENCE.** The compression run that produced it yielded a
+**degenerate model**: probe MMLU (limit 40/subtask, n=2280, seed 1234, fp16) = **23.42%**,
+below the 25% chance level, and a pre-heal health gate initially refused it with KL=NaN.
+Even 60M tokens of KD-healing could not restore capability (26.27% — see
+`../heal/llama2-7b-base/README.md`).
+
+Archived as a documented negative result. The ASVD chat instance
+(`asvd_llm/unheal/llama2-7b/`) is healthy; the failure is specific to this base run
+(abs_mean scaling, alpha=0.5, seed=233, n_calib=256, ppl sensitivity, ratio target 0.589).
+
+- Factored params: **4,020,677,888** (factored 3,532,823,808 + untouched 487,854,080;
+  budget 4.0231B, matched to the TR-3 llama2-7b row)
+
+Load with `lowrank_linear.py` in this dir (identical recipe to the sibling dirs).
+Compression provenance is in `provenance.json`.