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

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

@@ -0,0 +1,65 @@
+---
+license: llama2
+base_model: NousResearch/Llama-2-7b-hf
+tags:
+  - asvd
+  - low-rank
+  - model-compression
+  - llama2
+  - negative-result
+---
+
+# ⚠️ ASVD: LLaMA-2-7B base @ 4.02B, KD-healed — FAILED ARTIFACT (archived negative result)
+
+**DO NOT USE THIS CHECKPOINT FOR INFERENCE.** It is archived for reproducibility as a
+documented failure, not released as a working model.
+
+## What happened
+
+ASVD (official-style: activation-aware SVD with abs_mean scaling, ppl-based sensitivity,
+alpha=0.5, seed=233, 256 calibration samples) compressed LLaMA-2-7B base to **4,020,677,888**
+params. The compressed model was **capability-dead on arrival** (forward ≈ random; a first
+heal attempt was auto-refused by a pre-training health gate with KL=NaN). A subsequent heal
+run (60M tokens, pure forward-KL distillation from the dense 7B base teacher, identical
+recipe to the healthy sibling dirs) completed but **failed to restore any capability**:
+
+| Signal (12 milestone probes) | start | 60M |
+|---|---:|---:|
+| held-out KL (KD-domain imitation) | 7.2970 | 0.2259 |
+| **probe MMLU (limit 40/subtask, n=2280, seed 1234, fp16)** | 23.42% | **26.27%** |
+
+KL collapsed 33x while MMLU stayed pinned at chance (25%) for the entire run — imitation on
+the KD corpus recovered, knowledge did not.
+
+Same healing recipe, same budget — healthy siblings:
+
+| Checkpoint (this repo) | probe MMLU after heal |
+|---|---:|
+| **ASVD base 7B (this dir)** | **26.27% (dead)** |
+| SVD-LLM V1 base 7B (`svd_llm/heal/llama2-7b-base/`) | 34.12% |
+| SVD-LLM V2 base 7B (`svd_llm2/heal/llama2-7b-base/`) | 35.66% |
+| TR-3 base 7B (`tr3/heal/llama2-7b-base/`) | 41.97% |
+| FP16 dense 7B base teacher (same probe) | 47.41% |
+
+Chat-line ASVD (`asvd_llm/heal/llama2-7b/`) works, so the failure is specific to this base
+instance, not the method as a whole.
+
+## How to load (forensics only)
+
+```python
+import json
+from lowrank_linear import build, materialize_all
+
+model = build("NousResearch/Llama-2-7b-hf",
+              json.load(open("ranks.json")), "state_dict_fp16.pt",
+              device="cuda")
+materialize_all(model)
+```
+
+Full healing provenance (trajectory of all 12 probes, SHA256 of every input) is in
+`provenance.json`. Compression provenance is in the sibling `unheal` dir.
+
+## License
+
+Derived from Meta's LLaMA 2 (`NousResearch/Llama-2-7b-hf`). Subject to the
+**LLaMA 2 Community License** — see https://ai.meta.com/llama/license/.