Upload svd_llm2/unheal/llama2-7b-base/provenance.json with huggingface_hub

svd_llm2/unheal/llama2-7b-base/provenance.json

@@ -0,0 +1,21 @@
+{
+  "method": "SVD-LLM V2 (NAACL 2025): Algorithm 1 heterogeneous ratio allocation + Algorithm 2 loss-optimized truncation; official code unreleased, faithful reimplementation (project IMPLEMENTATION_NOTES)",
+  "source_model": "NousResearch/Llama-2-7b-hf",
+  "source_revision": "8efe6c9b93655b934e27bd9981e3ec13e55aee9d",
+  "params_total": 4023134464,
+  "param_budget_target": "4.0231B, matched to the TR-3 llama2-7b(-base) row",
+  "overall_keep_ratio": 0.5680,
+  "compression_artifacts": {
+    "compressed_checkpoint_tr3fmt": "checkpoints/svdllm2_base_7b/tr3fmt/state_dict_fp16.pt",
+    "compressed_checkpoint_tr3fmt_sha256": "0d4448da4e196a60a1c0086908fec20755a2fe629d42d4fdfc20c5a8c2057f1a",
+    "ranks_json_sha256": "cbf5384f256345afac09962a58e2979b1ff2f3b63f4f7e5055b61c6f30554ed4",
+    "v2_allocation_files": "checkpoints/svdllm2_base_7b/{v2_ranks.json, v2_lmin.json}",
+    "raw_svd_output": "checkpoints/svdllm2_base_7b/NousResearch_Llama_2_7b_hf_svdllmv2_keep_0.5680.pt",
+    "rawgram_cache": "checkpoints/svdllm2_base_7b/NousResearch_Llama_2_7b_hf_rawgram_wikitext2_256_3.pt"
+  },
+  "sha_source": "checkpoint sha256 recorded by the sibling KD-heal run's input-identity gate (job 93283)",
+  "probe_mmlu_unhealed": 0.24342105263157895,
+  "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 93283). 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)"
+}