Upload svd_llm/unheal/llama2-13b-base/provenance.json with huggingface_hub

svd_llm/unheal/llama2-13b-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 ~41.6% removal across all 7 projection types of all 40 layers",
+  "source_model": "NousResearch/Llama-2-13b-hf",
+  "source_revision": "4b54e9fdd2ebd8db0901d37b98d3a53bbfaa4503",
+  "params_total": 7737799680,
+  "param_budget_target": "7.7403B, matched to the TR-3 llama2-13b(-base) row",
+  "overall_keep_ratio": 0.5842,
+  "compression_artifacts": {
+    "compressed_checkpoint_tr3fmt": "checkpoints/svdllm1_base_13b/tr3fmt/state_dict_fp16.pt",
+    "compressed_checkpoint_tr3fmt_sha256": "bcf2f34e1c7ba25c796057ee555bb3863d65e88f8f75fef35ca4477583901efc",
+    "ranks_json_sha256": "5c9ed0db91c9c81c51e62e6f82c7c2772478cc353c6f28040164843e64219b55",
+    "raw_svd_output": "checkpoints/svdllm1_base_13b/NousResearch_Llama_2_13b_hf_whitening_then_update_0.5842.pt",
+    "profiling_cache": "checkpoints/svdllm1_base_13b/NousResearch_Llama_2_13b_hf_profiling_wikitext2_256_3.pt"
+  },
+  "sha_source": "checkpoint sha256 recorded by the sibling KD-heal run's input-identity gate (job 93281)",
+  "probe_mmlu_unhealed": 0.2574561403508772,
+  "probe_mmlu_teacher_same_harness": 0.5425438596491228,
+  "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 93281). Full-benchmark eval intentionally not run on base models.",
+  "eval_protocol": "eval_protocol_llama2_13b_base.json"
+}