Upload svd_llm2/unheal/llama2-13b-base/provenance.json with huggingface_hub

svd_llm2/unheal/llama2-13b-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-2-13b-hf",
+  "source_revision": "4b54e9fdd2ebd8db0901d37b98d3a53bbfaa4503",
+  "params_total": 7740277760,
+  "param_budget_target": "7.7403B, matched to the TR-3 llama2-13b(-base) row",
+  "overall_keep_ratio": 0.5740,
+  "compression_artifacts": {
+    "compressed_checkpoint_tr3fmt": "checkpoints/svdllm2_base_13b/tr3fmt/state_dict_fp16.pt",
+    "compressed_checkpoint_tr3fmt_sha256": "fb5a1342778d945a642cd1cb1fb74f7121e7210877e76e1b1fd6e354c7a77ffc",
+    "ranks_json_sha256": "3ea3ed43101dffde74715cb68b16ea181b035c35074b15365e35557f88f7d211",
+    "v2_allocation_files": "checkpoints/svdllm2_base_13b/{v2_ranks.json, v2_lmin.json}",
+    "raw_svd_output": "checkpoints/svdllm2_base_13b/NousResearch_Llama_2_13b_hf_svdllmv2_keep_0.5740.pt",
+    "rawgram_cache": "checkpoints/svdllm2_base_13b/NousResearch_Llama_2_13b_hf_rawgram_wikitext2_256_3.pt"
+  },
+  "sha_source": "checkpoint sha256 recorded by the sibling KD-heal run's input-identity gate (job 93338)",
+  "probe_mmlu_unhealed": 0.2614035087719298,
+  "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 93338). Full-benchmark eval intentionally not run on base models.",
+  "eval_protocol": "eval_protocol_llama2_13b_base.json"
+}