HectorHHZ · HectorHHZ · Sep 21, 2026 · Sep 21, 2026
diff --git a/docs/experiments/artifacts/gpt-oss-120b-gpqa-main-calib64/README.md b/docs/experiments/artifacts/gpt-oss-120b-gpqa-main-calib64/README.md
@@ -0,0 +1,47 @@
+# GPT-OSS-120B held-out GPQA-main artifacts
+
+This directory contains the public, question-safe artifacts behind the
+GPT-OSS-120B base and IntDim-E/L/G 50% results in the parent experiment record.
+
+## Dataset
+
+The access-controlled Hugging Face dataset
+[`jayzou3773/less-is-moe-gpqa-main-calibration-64`](https://huggingface.co/datasets/jayzou3773/less-is-moe-gpqa-main-calibration-64),
+revision `7134dfef5af4605eae0706c30efa9226f49aed96`, contains the exact 64-row
+calibration set and 384-row held-out evaluation set. It comes from
+`Idavidrein/gpqa`, config `gpqa_main`, revision
+`633f5ee89ab8ad4522a9f850766b73f62147ffdd`.
+
+GPQA asks users not to reveal examples online. This public repository therefore
+commits only `dataset/selection-manifest.json` and `dataset/split-manifest.json`;
+these contain provenance, selection rules, UUIDs, and hashes without question
+text.
+
+```text
+calibration.jsonl SHA-256: 99e61c8c1bd4321162f1f3895d62e889b9a8b11be9627064207effce70e324c7
+test.jsonl SHA-256:        c7475f3382e65a23a5f3c1da97947e1ac9c110686efcc23df16c44c9b65c44cb
+selection manifest:       fd4bb8cd9be999a6ad780c14570af17fdd434b63da2199f12549f73aaf8b8f4d
+split manifest:           ff0bf0307c875869ea9b90feab29413a9b6a7777ed844c5add4626b93d7b941a
+```
+
+## Results
+
+Each result directory contains:
+
+- `predictions.sanitized.jsonl`: all 3,072 per-sample decisions, with question
+  text, options, answer text, raw generation, and final-answer text removed;
+- `metrics.json`: aggregate, repeat-level, and hierarchy metrics from the frozen
+  `graded-v4` scorer;
+- `run.json`: model, runtime, prompt, sampling, hashes, and environment settings;
+- `grading-audit.json`: raw-prediction hashes and the uniform regrading audit.
+
+| Run | Correct | Total | Accuracy | Unparsed | Completion tokens |
+|---|---:|---:|---:|---:|---:|
+| Base BF16 | 2,365 | 3,072 | 76.99% | 3 | 28,039,222 |
+| IntDim-E 50% | 779 | 3,072 | 25.36% | 188 | 1,586,117 |
+| IntDim-L 50% | 1,935 | 3,072 | 62.99% | 12 | 15,567,420 |
+| IntDim-G 50% | 1,796 | 3,072 | 58.46% | 44 | 40,743,679 |
+
+Summing `correct`, `completion_tokens`, and null `prediction` fields in the
+sanitized files reproduces these metrics. `artifact-manifest.json` records the
+SHA-256 of every committed result and both private raw/graded prediction files.
diff --git a/docs/experiments/artifacts/gpt-oss-120b-gpqa-main-calib64/artifact-manifest.json b/docs/experiments/artifacts/gpt-oss-120b-gpqa-main-calib64/artifact-manifest.json
@@ -0,0 +1,112 @@
+{
+  "schema_version": 1,
+  "dataset": {
+    "name": "Idavidrein/gpqa",
+    "config": "gpqa_main",
+    "revision": "633f5ee89ab8ad4522a9f850766b73f62147ffdd",
+    "hf_dataset": "jayzou3773/less-is-moe-gpqa-main-calibration-64",
+    "hf_revision": "7134dfef5af4605eae0706c30efa9226f49aed96",
+    "visibility": "private/access-controlled",
+    "source_count": 448,
+    "calibration_count": 64,
+    "evaluation_count": 384,
+    "calibration_jsonl_sha256": "99e61c8c1bd4321162f1f3895d62e889b9a8b11be9627064207effce70e324c7",
+    "evaluation_jsonl_sha256": "c7475f3382e65a23a5f3c1da97947e1ac9c110686efcc23df16c44c9b65c44cb",
+    "selection_manifest_path": "dataset/selection-manifest.json",
+    "selection_manifest_sha256": "fd4bb8cd9be999a6ad780c14570af17fdd434b63da2199f12549f73aaf8b8f4d",
+    "split_manifest_path": "dataset/split-manifest.json",
+    "split_manifest_sha256": "ff0bf0307c875869ea9b90feab29413a9b6a7777ed844c5add4626b93d7b941a",
+    "question_text_committed": false,
+    "question_text_policy": "GPQA access terms request that examples not be revealed online."
+  },
+  "prediction_schema": {
+    "sanitized": true,
+    "omitted_fields": [
+      "question",
+      "options",
+      "answer",
+      "raw_output",
+      "final_answer"
+    ],
+    "question_identifier": "question_sha256",
+    "rows_per_run": 3072
+  },
+  "results": {
+    "base": {
+      "predictions_path": "results/base/predictions.sanitized.jsonl",
+      "predictions_sha256": "57d161f995af8bbea2504a6e20077a5ef9a3b0e75ccdce90ebbe78e31991729f",
+      "metrics_path": "results/base/metrics.json",
+      "metrics_sha256": "9c518fec8b89092ff470ffaffe8eb744d37800eecaf097f636260760e167c1ff",
+      "run_path": "results/base/run.json",
+      "run_sha256": "e0ed3f3711ae3d707b74bbf1a3734ca1ec79634783d13902412240a3241b9c58",
+      "grading_audit_path": "results/base/grading-audit.json",
+      "grading_audit_sha256": "469cb3c49e42c86446b1cfe40f090100b3a7b19875d51d1cf027785093571f10",
+      "source_raw_predictions_sha256": "cf1b0510f5ca2d2d09876904c9616b00619955ce98e57724dbcb492521962605",
+      "source_graded_predictions_sha256": "e9fea158716ccd9a9723223bff07405bdac03ee1c24ecb91fdf1806622cb876f",
+      "rows": 3072,
+      "correct": 2365,
+      "total": 3072,
+      "accuracy": 0.7698567708333334,
+      "unparsed": 3,
+      "truncated": 0,
+      "completion_tokens": 28039222
+    },
+    "intdim-e-50": {
+      "predictions_path": "results/intdim-e-50/predictions.sanitized.jsonl",
+      "predictions_sha256": "e68ae458f24875d8c59d79007c6df9744ad2776b4158aa330221c1d94d08c81a",
+      "metrics_path": "results/intdim-e-50/metrics.json",
+      "metrics_sha256": "0b1afd1cd3f2f6cf4e4aae9f194a4ca082a2702e64b875b25acbc17f37a3ceba",
+      "run_path": "results/intdim-e-50/run.json",
+      "run_sha256": "132358277d876400a844cf1bc6ca0fddf7aa337da516d45cbd1ead6d94650de7",
+      "grading_audit_path": "results/intdim-e-50/grading-audit.json",
+      "grading_audit_sha256": "c091ac57c6fe9dd005f0cd74713a33871e5d23c0dcf33c662da0f274d601b3ba",
+      "source_raw_predictions_sha256": "9e8f315543e692e8fb0e8b570d2a72b3ae5845f7cb51e8e97ec3974ed172d7de",
+      "source_graded_predictions_sha256": "abf11daf5087340f5183c16ad5f079d03b94b3cff0dac7cfa003f279dbc0ee13",
+      "rows": 3072,
+      "correct": 779,
+      "total": 3072,
+      "accuracy": 0.2535807291666667,
+      "unparsed": 188,
+      "truncated": 0,
+      "completion_tokens": 1586117
+    },
+    "intdim-l-50": {
+      "predictions_path": "results/intdim-l-50/predictions.sanitized.jsonl",
+      "predictions_sha256": "5e35af594b4ddaac2f0a7cb2b31404044a590427c3c297841ff8637450c0ae0c",
+      "metrics_path": "results/intdim-l-50/metrics.json",
+      "metrics_sha256": "e94000bfa29311c0c099aab077e637135c4c78afa7ed182def1f29a6569c3cb8",
+      "run_path": "results/intdim-l-50/run.json",
+      "run_sha256": "977df904b11aaac0bf597580d6b9d4a34896ce96985887940a1cd22144366527",
+      "grading_audit_path": "results/intdim-l-50/grading-audit.json",
+      "grading_audit_sha256": "a17701d9068815cf4b37ddeff39d86f8406192c05e9e7d9d55770e1b8203acd8",
+      "source_raw_predictions_sha256": "8c8d1b7696130a0f7d7b621d181871632827084ba4544f43fc2ee41106e15934",
+      "source_graded_predictions_sha256": "beb5edae9ae47812eae4c6b179797e00cd14e30d0c990f5dfc3d5026bf0d8df4",
+      "rows": 3072,
+      "correct": 1935,
+      "total": 3072,
+      "accuracy": 0.6298828125,
+      "unparsed": 12,
+      "truncated": 0,
+      "completion_tokens": 15567420
+    },
+    "intdim-g-50": {
+      "predictions_path": "results/intdim-g-50/predictions.sanitized.jsonl",
+      "predictions_sha256": "6e076e1ff9c44dd244692725733fa742719fc8ec87ae8cb1f9f80beffb742e9d",
+      "metrics_path": "results/intdim-g-50/metrics.json",
+      "metrics_sha256": "a558d3f1a7d5cf4fb69a4caca4c25ff40c76b099d88a5f3883367090c100054a",
+      "run_path": "results/intdim-g-50/run.json",
+      "run_sha256": "f843bb04e8922e0deb56be9ba9645a36364ccb90e1bb907cac016b7e63bfe997",
+      "grading_audit_path": "results/intdim-g-50/grading-audit.json",
+      "grading_audit_sha256": "ff6342db99633e36cf448864e743f5f81aefe7207ef87dc1650eb852f0fa825f",
+      "source_raw_predictions_sha256": "c6c001ec04721b052f11bd7d6f8bfaae17820645ef3ba255a3c2a9698e9093c7",
+      "source_graded_predictions_sha256": "1401424a393af099eefea993e0de21a1434d7e4ed6bfda44cb4d41a9eb417907",
+      "rows": 3072,
+      "correct": 1796,
+      "total": 3072,
+      "accuracy": 0.5846354166666666,
+      "unparsed": 44,
+      "truncated": 0,
+      "completion_tokens": 40743679
+    }
+  }
+}
diff --git a/docs/experiments/artifacts/gpt-oss-120b-gpqa-main-calib64/dataset/selection-manifest.json b/docs/experiments/artifacts/gpt-oss-120b-gpqa-main-calib64/dataset/selection-manifest.json
@@ -0,0 +1,124 @@
+{
+  "dataset": "Idavidrein/gpqa",
+  "dataset_config": "gpqa_main",
+  "dataset_split": "train",
+  "dataset_revision": "633f5ee89ab8ad4522a9f850766b73f62147ffdd",
+  "source_file": "gpqa_main.csv",
+  "source_file_sha256": "acdeeac8f622267f2cd727d7d474202ea08dec80f7d3c3593b3ef8644f19b8e3",
+  "source_count": 448,
+  "samples": 64,
+  "test_samples": 384,
+  "split_policy": "64 deterministic calibration rows; all remaining canonical questions form test",
+  "selection_seed": 1234,
+  "selection": "numpy.random.default_rng(1234).permutation(448), first 64 rows",
+  "source_indices": [
+    343,
+    236,
+    382,
+    226,
+    47,
+    163,
+    190,
+    400,
+    370,
+    333,
+    436,
+    34,
+    192,
+    336,
+    232,
+    199,
+    426,
+    85,
+    357,
+    5,
+    283,
+    301,
+    254,
+    425,
+    217,
+    141,
+    258,
+    285,
+    105,
+    378,
+    392,
+    150,
+    193,
+    257,
+    415,
+    312,
+    324,
+    288,
+    57,
+    281,
+    97,
+    270,
+    328,
+    397,
+    315,
+    21,
+    434,
+    411,
+    327,
+    262,
+    207,
+    208,
+    423,
+    212,
+    126,
+    238,
+    435,
+    130,
+    268,
+    359,
+    444,
+    329,
+    266,
+    279
+  ],
+  "source_indices_sha256": "790c4c22309def44542965fdde7c5f38f1d8e354602640cfb31518134b8d92e6",
+  "option_seed": 42,
+  "option_order": "per-record deterministic SHA-256 seed; prevents correct-answer position bias",
+  "reasoning_trace_field": "Explanation",
+  "answer_field": "Correct Answer",
+  "reasoning_trace_kind": "canonical expert-written answer rationale; not a model-generated hidden chain-of-thought",
+  "messages": "user = canonical Question plus shuffled choices; assistant = Explanation plus Answer: <letter>",
+  "loader_text": "messages[].content concatenated with one space; no chat template",
+  "truncation": false,
+  "all_rows_explanation_nonempty": 448,
+  "all_rows_explanation_chars": {
+    "minimum": 99,
+    "median": 645.5,
+    "maximum": 3225
+  },
+  "selected_explanation_chars": {
+    "minimum": 164,
+    "median": 646.0,
+    "maximum": 3225,
+    "total": 51977
+  },
+  "selected_text_chars": {
+    "minimum": 549,
+    "median": 1356.0,
+    "maximum": 5064,
+    "total": 98282
+  },
+  "selected_domain_counts": {
+    "Biology": 10,
+    "Chemistry": 25,
+    "Physics": 29
+  },
+  "selected_subdomain_count": 13,
+  "selected_answer_letter_counts": {
+    "A": 14,
+    "B": 19,
+    "C": 15,
+    "D": 16
+  },
+  "calibration_file": "calibration.jsonl",
+  "calibration_file_sha256": "99e61c8c1bd4321162f1f3895d62e889b9a8b11be9627064207effce70e324c7",
+  "test_file": "test.jsonl",
+  "test_file_sha256": "c7475f3382e65a23a5f3c1da97947e1ac9c110686efcc23df16c44c9b65c44cb",
+  "publication": "private local artifact; GPQA access terms prohibit revealing examples online"
+}