Upload svd_llm/unheal/llama2-7b-base/provenance.json with huggingface_hub

svd_llm/unheal/llama2-7b-base/provenance.json

@@ -0,0 +1,20 @@
+{
+  "method": "SVD-LLM V1 (ICLR 2025): truncation-aware data whitening (256 wikitext2 samples) + closed-form parameter update; official code AIoT-MLSys-Lab/SVD-LLM @ 7538cca, uniform 41.93% removal across all 7 projection types of all 32 layers",
+  "source_model": "NousResearch/Llama-2-7b-hf",
+  "source_revision": "8efe6c9b93655b934e27bd9981e3ec13e55aee9d",
+  "params_total": 4021989376,
+  "param_budget_target": "4.0231B, matched to the TR-3 llama2-7b(-base) row",
+  "overall_keep_ratio": 0.5807,
+  "compression_artifacts": {
+    "compressed_checkpoint_tr3fmt": "checkpoints/svdllm1_base_7b/tr3fmt/state_dict_fp16.pt",
+    "compressed_checkpoint_tr3fmt_sha256": "1009aa0e0a873b4e4957ffd9df855027ff0f695ab2ef0f0ac7b0f4fd91295a69",
+    "ranks_json_sha256": "715d09b26b58bbc4052eb425dc041e5d8b3acb5cc6cef729e2ca34c76c0e2943",
+    "raw_svd_output": "checkpoints/svdllm1_base_7b/NousResearch_Llama_2_7b_hf_whitening_then_update_0.5807.pt",
+    "profiling_cache": "checkpoints/svdllm1_base_7b/NousResearch_Llama_2_7b_hf_profiling_wikitext2_256_3.pt"
+  },
+  "sha_source": "checkpoint sha256 recorded by the sibling KD-heal run's input-identity gate (job 93282)",
+  "probe_mmlu_unhealed": 0.256140350877193,
+  "probe_mmlu_teacher_same_harness": 0.47412280701754383,
+  "probe_note": "lm-eval 0.4.12, MMLU limit=40 questions/subtask, n=2280, seed 1234, fp16; measured as the step-0 eval of the sibling KD-heal run (RIKYU job 93282). Full-benchmark eval intentionally not run on base models.",
+  "eval_protocol": "eval_protocol_llama2_7b_base.json (mirrored at tr3/heal/llama2-7b-base/ in this repo)"
+}