| { |
| "repetition": "r1", |
| "settings": { |
| "study": "v7-r1", |
| "class_labels": [ |
| "TRUE", |
| "FALSE", |
| "UNKNOWN" |
| ], |
| "train_old_families": 4032, |
| "train_multi_groups_per_size": 84, |
| "auxiliary_old_families": 168, |
| "auxiliary_multi_groups_per_size": 7, |
| "test_old_families": 300, |
| "test_multi_groups_per_size": 40, |
| "transfer_old_families": 100, |
| "transfer_multi_groups_per_size": 12, |
| "max_attempts_per_family": 200, |
| "max_length": 512, |
| "epochs": 3, |
| "effective_batch": 32, |
| "microbatch": 8, |
| "evaluation_batch": 32, |
| "backbone_lr": 2e-05, |
| "head_lr": 0.0001, |
| "weight_decay": 0.01, |
| "gradient_clip": 1.0, |
| "warmup_fraction": 0.05, |
| "candidate_steps": [ |
| 0, |
| 250, |
| 600, |
| 900, |
| 1134 |
| ], |
| "chunk_steps": 378, |
| "max_training_calls_per_arm": 3, |
| "temperature_min": 0.05, |
| "temperature_max": 1000.0, |
| "temperature_iterations": 64, |
| "ece_bins": 10, |
| "high_confidence_threshold": 0.9, |
| "bootstrap_repetitions": 2000, |
| "bootstrap_confidence": 0.95, |
| "noninferiority_margin": 0.02, |
| "cpu_threads": 2, |
| "gpu": "L4", |
| "cpu_cores": 4, |
| "memory_mib": 32768, |
| "pilot_timeout_seconds": 420, |
| "training_timeout_seconds": 1050, |
| "selection_timeout_seconds": 600, |
| "test_timeout_seconds": 900, |
| "retry_timeout_seconds": 1050, |
| "max_retry_calls": 1, |
| "startup_timeout_seconds": 180, |
| "scaledown_seconds": 2, |
| "gpu_hourly_usd": 0.8, |
| "cpu_core_hourly_usd": 0.0473, |
| "memory_gib_hourly_usd": 0.008, |
| "compute_allocation_usd": 4.0, |
| "storage_reserve_usd": 0.5, |
| "profile": "amandogra2016", |
| "volume": "decision-model-lab-v7-runs", |
| "app_name": "decision-model-lab-v7-replication", |
| "hf_namespace": "axondendriteplus", |
| "parent_repo": "axondendriteplus/decision-lab-v4-smollm2-360m-mixed-unknown", |
| "parent_commit": "c4ee886cd3a718036297edb56b0202b590457cf2", |
| "parent_directory": "artifacts/v4-releases/treatment", |
| "dataset_seed": 731, |
| "model_seed": 731, |
| "bootstrap_seed": 731 |
| }, |
| "budget": { |
| "hourly_requested_usd": 1.2452, |
| "call_ceiling_usd": { |
| "pilot": 0.2082251111111111, |
| "training": 0.42613511111111113, |
| "selection": 0.2704851111111111, |
| "test": 0.37425177777777785, |
| "retry": 0.42613511111111113 |
| }, |
| "maximum_compute_usd": 3.835907777777778 |
| }, |
| "code_files": { |
| "decision_lab/release.py": "148cd56fc33111a719582301e551534d8d4bbcc04ba31aa22993935c5f172c70", |
| "decision_lab/config.py": "29b30051d14b684cad639f1509243fcf709f3ea39bfa002876e0726c4d1b1be1", |
| "decision_lab/model_config.py": "f4781f6881239170498fd4fedd06d2db59fde992969325fceca6064585395c2b", |
| "decision_lab/baseline.py": "dd71a9bb85d153abeeab20acdb4b612fbba17f30cfd35e6a1d0df6c8559d127e", |
| "decision_lab/budget.py": "482868f106e4477739d7b1cc1834ad7429b065f5651bbedd86eaa3b11718aa96", |
| "decision_lab/__init__.py": "91447944015cec709e8aa7655f7e9d64e1e4508e7023a57fe3746911c0fc6fed", |
| "decision_lab/tokenization.py": "8e3097fe659615fc380acf05438c1d5d510515b386d5f55afaf2b06995ec600a", |
| "decision_lab/cli.py": "82288bfa6c7e3d5d57f2a5fd77c20da1467cf2a8f5568ca6c23c4476ce61a03d", |
| "decision_lab/publication.py": "264743e4f53186e0ba915d821c6d4e3a6b70ac1123c3ba933fc87e48372cb751", |
| "decision_lab/__main__.py": "b2cabffd6debed2e83f56947a2b4b799c4172886bca562077a7d6407db671f11", |
| "decision_lab/training/pilot.py": "402613126a5afe6f9e31cbbd54bf8225bdc5a2801606838a2ad05742dc74fa8e", |
| "decision_lab/training/runner.py": "45d4d224fb2a5d1975df8cc2630b44c4a05f7bb55567b029f367e3d689701182", |
| "decision_lab/training/config.py": "e37b9187f57ea42afa645cd1a454684937c396718266e4d470454f6c03710a81", |
| "decision_lab/training/checkpoint.py": "4a9524ff15d0752a7171066309d712db41817506efa28eaf8263c9f02b730878", |
| "decision_lab/training/protocol.py": "6156e7abe2c8f48b869d3fb43623af27de1c932db35eddb695a484dc1bc7fef6", |
| "decision_lab/training/ledger.py": "8281e20c048974d6281b899866f524b4ee414dbcbcfcf5cce8fedd2c87719350", |
| "decision_lab/training/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", |
| "decision_lab/training/core.py": "a1e42e2bd2fa0c4412390592501ff6540a798611a07a790eb7ac1e565fdb0b6e", |
| "decision_lab/training/gate.py": "37a6663998d7dc146f4ac0442921ee96f4c797fcb4fa94aa1ac4b325ed99c9a9", |
| "decision_lab/models/decision_head.py": "9fd465586b90bcad6b36d050751b2f8b4c9ea6201476573c7516588f10bcdf3d", |
| "decision_lab/models/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", |
| "decision_lab/models/token_model.py": "106ce581e428e4cc5e4ffa8e061bc3468dd8ed70623188d7d5e2d5e94bf9dbc7", |
| "decision_lab/models/scratch_transformer.py": "e16551ee7f9117c240faba489594a4643a5a0246215938d4693b6618e7fe98d8", |
| "decision_lab/v2/recovery.py": "8b5f8e2daaa8963e1cbc377e4c48ed84c724061f5e1613e06a97c497c900e1ac", |
| "decision_lab/v2/budget.py": "ffe3c39f8e9a1435db36fcf2edb4df92f4900a2748cf9a13b5d6874779b8ee70", |
| "decision_lab/v2/protocol.py": "47b09ffbc2f7d563d5102757058ac623721c68b1df9d5c0dd718fe39f2daf804", |
| "decision_lab/v2/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", |
| "decision_lab/v2/report.py": "0f3696164c4e06a5b7218396ef25ca7d5f5d886b3db5353237b8cdfdb9ffabd6", |
| "decision_lab/v2/data.py": "1f50f8ed64ff9352b5ce11c004eb8bfa6d370d3af5cd6af40cef05b661bd1cd1", |
| "decision_lab/v2/study.py": "a0648f80cc063d71451f98c108a7a163bb4daec6608d342b4d5aeaddc638162a", |
| "decision_lab/v3/protocol.py": "a2701d627064b58acce4aca752b035dadf0fd88c4cba97a46cb8511945ac7766", |
| "decision_lab/v3/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", |
| "decision_lab/v3/selection.py": "ceb0cfeed5e765573b4ddd653a4d213e130d5ea92b206617fba3128c04bd7421", |
| "decision_lab/v3/data.py": "197b339daf9a8013e4f03f12f4e6133dfc4e43423c0ef2711ea8b98facd27dc1", |
| "decision_lab/v3/study.py": "6cc67fef72e47792de9ebf4bb34fa6727872a683b536f393cd3dd4184fc39717", |
| "decision_lab/evaluation/metrics.py": "647619de2e0ad80b8701fc995f0b8e5a7fed67a01b3266000a86bf4bbd3d07b4", |
| "decision_lab/evaluation/baselines.py": "5a422bbc013a211633adad4b2c5cdc50c8d41862dccc532d2e15ff9ff8c6f241", |
| "decision_lab/evaluation/config.py": "7d8f82812f20b48851c239a22763c87947661e38822751a2c0c7c180767b28ee", |
| "decision_lab/evaluation/manifest.py": "016af5a65d5418e9b16e8c95387e93894adebde3f2e3a9ab050c2940d37fda85", |
| "decision_lab/evaluation/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", |
| "decision_lab/evaluation/export.py": "4f6ceec9eecdbd1d9950e1eadcaa58feff30c9c1f9fe8f78f5bfb58992b4c0f3", |
| "decision_lab/evaluation/robustness.py": "a8c7c7e5229f55986edbc15bb7355548fb6928fe7099484ad71e5be12e978018", |
| "decision_lab/evaluation/learning.py": "870c1d08489c12df527860137c3d2eea07ca1ffedbb346bbc321cf2a8e94f7ad", |
| "decision_lab/evaluation/performance.py": "8d5e919f57015c739217ffacf95e801e7345275d2654306e73ab4572246d4016", |
| "decision_lab/evaluation/report.py": "f34af511ebfb21d0c1cbe58e67b6ae2dd4daefc1cf6c484245d36f65481059ac", |
| "decision_lab/evaluation/study.py": "20dbbfe533d127612ca597df0c04b1d0df505c8c919a88dccd272f4e3215f743", |
| "decision_lab/data/render.py": "07378b4a7243666e4399be7000cd3d8e67ad0b5374b195aaf7742596567adcb0", |
| "decision_lab/data/generate.py": "5f9ae68bef5f416b76c737b1c8b7f0858bd82d508e0986abf1ea0f2dd9002f89", |
| "decision_lab/data/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", |
| "decision_lab/data/validate.py": "4159475118d38acbfdcd7c082b1769bf957e61c3769cdaadac0f099f1009981a", |
| "decision_lab/data/split.py": "1a7c3f5b81df2f5503fcfeabd916129706171f79a7e77c833cb9a604c76d7dfe", |
| "decision_lab/data/solve.py": "d7d520c5843d578841149f02cd01fa497178bca07c6f4d2445976587265b7ab2", |
| "decision_lab/data/schema.py": "e54c5df39c2659e29f1c1a03361d80f0f99a2058e3fed0924aff8fed276c3aaa", |
| "v4_lab/protocol.py": "9f97eba0d2a45a61e04a149de87d2c6192134e50b0e3bb45373167485d209df0", |
| "v4_lab/__init__.py": "01ba4719c80b6fe911b091a7c05124b64eeece964e09c058ef8f9805daca546b", |
| "v4_lab/data.py": "7a6d4c34cfacefc8c1a6593dc30168681df110b529799bac8217aba11dc4ef54", |
| "v4_lab/study.py": "79860f762b946299f6d7ab5ee1aab214969b5dcb95eb0c559caa7fd4255bed40", |
| "v5_lab/analysis.py": "e5b501730bd7cd2b023aa42462b0910a7b53381cd653e224338066c2262c7eda", |
| "v5_lab/protocol.py": "d5b99e48253f172c0c5b86bb9c205bf63497ca99721489ca2dd4ac1f15f7ff7d", |
| "v5_lab/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", |
| "v5_lab/data.py": "16be48f335775c39027dc31c2f171c2e8605fc994f0d1fccb1e5004b7d382504", |
| "v5_lab/study.py": "6f9d4803dcd53bfcf531459d71ed69386bab9a95be736129b28ca7d13aba0678", |
| "v6_lab/analysis.py": "f3d6d505e2a4824fe5b7a7255d4cd0564cc65a666e28a6d1bab668dff217c598", |
| "v6_lab/protocol.py": "b4e2c726b55b92d9a8dca103950527d6a590d9d76ed6a21fdcae5c4a2e5b8b63", |
| "v6_lab/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", |
| "v6_lab/data.py": "4ea6afa232dfb59d4db726271adc5af1315647f1d80734abc4db4e667b3ca456", |
| "v6_lab/study.py": "5403a78263ea3edb877030f9659168bccbab00c7fa327f93ca9e93239eb9f4b4", |
| "v7_lab/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", |
| "v7_lab/analysis.py": "78462886f7c377f108fdbb0fe55b425dc7e063a00d6bb9b0468693bb7e733cc3", |
| "v7_lab/data.py": "381a87b6cd8aff4a9d053f0806c05e9f872d227d49c7be312c57f9d43fc366cb", |
| "v7_lab/protocol.py": "df70db7f6fe233d9e9388d3f416e3de63bd8cb7464401dc4ad87424e7d905bf6", |
| "v7_lab/study.py": "9885dd04722c8fec585d38b921ec44e32698e8e17efa866a1f63bbf734437c6c" |
| }, |
| "parent_files": { |
| "LICENSE": "cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30", |
| "NOTICE": "dd232662e7484ce000e822044e188bda4a41419882586bdea3a955470e08f406", |
| "README.md": "d8b9c9d25ea395baae630710e519f423bb9e87969318dae9ff7b0cb33eb5eec5", |
| "artifact-manifest.json": "5a54b1dab9d5a7f2948acb5b996f2a2ff37b0db9a0ea7191fcab7d3a81c880d7", |
| "calibration.json": "d5214ed32b472446f6d7a215285803d1ab41aace3cdc82f2ae59ac9354f07331", |
| "code/decision_lab/__init__.py": "91447944015cec709e8aa7655f7e9d64e1e4508e7023a57fe3746911c0fc6fed", |
| "code/decision_lab/config.py": "29b30051d14b684cad639f1509243fcf709f3ea39bfa002876e0726c4d1b1be1", |
| "code/decision_lab/data/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", |
| "code/decision_lab/data/render.py": "07378b4a7243666e4399be7000cd3d8e67ad0b5374b195aaf7742596567adcb0", |
| "code/decision_lab/data/schema.py": "e54c5df39c2659e29f1c1a03361d80f0f99a2058e3fed0924aff8fed276c3aaa", |
| "code/decision_lab/model_config.py": "f4781f6881239170498fd4fedd06d2db59fde992969325fceca6064585395c2b", |
| "code/decision_lab/models/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", |
| "code/decision_lab/models/decision_head.py": "9fd465586b90bcad6b36d050751b2f8b4c9ea6201476573c7516588f10bcdf3d", |
| "code/decision_lab/models/scratch_transformer.py": "e16551ee7f9117c240faba489594a4643a5a0246215938d4693b6618e7fe98d8", |
| "code/decision_lab/models/token_model.py": "106ce581e428e4cc5e4ffa8e061bc3468dd8ed70623188d7d5e2d5e94bf9dbc7", |
| "code/decision_lab/release.py": "148cd56fc33111a719582301e551534d8d4bbcc04ba31aa22993935c5f172c70", |
| "code/decision_lab/training/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", |
| "code/decision_lab/training/checkpoint.py": "4a9524ff15d0752a7171066309d712db41817506efa28eaf8263c9f02b730878", |
| "config.json": "e35580a1cc1ead628fa640b7b725fa560593ac3cdcb37ce451e8b4cf1a898b7f", |
| "evaluation.json": "65b0e285d043c8a1465779e08f578c45470693ad13bdd8c0f9dc441a6e561f91", |
| "infer.py": "a51efb294777dfaa2a554ea5c0d1bd6b0631fed92fe846aa07dd23d38fb88a14", |
| "model.safetensors": "417f2b856d0b117d9c8e532539d026654d4e28f40b34f3ddf62e0bf3660cc7b1", |
| "prompt-template.txt": "098ead6f8ed63d3e5ebb0f8848ceb72a1d2fcbc74e01f5cc54d28e5325684874", |
| "protocol.json": "c94e5a6ba511d971f9a5e14f7fd315706a2641be3cce2c67ce6ef7a3c6fd5c0e", |
| "requirements.txt": "d0d672273f3284680e3d3453e8385cad0dde5aaf20f24adae22222d145fb6078", |
| "selection-freeze.json": "a7f724abc5bf55eb8946346636ca15e8cd5351325822531af72ee155e82d96dc", |
| "special_tokens_map.json": "79c83a585c13ce3020b4cfb5078ddc460d49c61dac2b569b37ae5833bb702f8c", |
| "tokenizer.json": "7d27c493c729a66ecefc837280b05d948b1ed50d130eebdbf911b1b36cf38ed7", |
| "tokenizer_config.json": "d41dfb95492c56d0ad9b9a44b4860a5ccbf445951ac0fed7fbf735826937d246", |
| "verification.json": "2e8382dc0fdda15699ad0cb8dcc0e29650c68868190c5cd4e660ea4b04f090f5" |
| }, |
| "data_files": { |
| "calibration.jsonl": "e6348f2d916990e4e55b0cef6f14ce0ed9cf6b6367e2e80334bc6f9149172b51", |
| "development.jsonl": "da1da907f50046e3e7af26d9d93d154b22bec43d94b249331887be7aa694141c", |
| "families.jsonl": "467f4ca69ddacf84a1818715442dc536bcb6a3d23877e0f691e16b584237afd2", |
| "manifest.json": "0fce69928e7f90ea0762e7fb9b570b8e03cee4cb1568a609e26a74010fa1c6ec", |
| "selection.jsonl": "65f448e09fc19c368d0f7c7d276a4326be13649a2b3ff3f57b0bd099867ea23d", |
| "selection_fit.jsonl": "afd8885da8e9fda188b207fb9ae2736ef940eb76b4a983999627f89da787ad2e", |
| "test_counterfactual.jsonl": "928cea632dd6975b847d46a5973cdb52af7c7d23f4bed28f2069e7746a8d97c2", |
| "test_depth.jsonl": "82f4c2560a1e9588668098b72e3501a86e7104895bbee0fcfccb7b2659a4d122", |
| "test_multi.jsonl": "06c178f129cc37d3ac4cdb17f2c4bf93f5370a8dd04078926c1ee7c4b01796e2", |
| "test_multi_depth.jsonl": "64e272d4dbafeb67f091156b0682000211d89ed1d808470414f17188cc654f99", |
| "test_multi_signed_chain.jsonl": "0e9f51a84c7d1a52e263e372918baffed373a6e1af3b775fc071a404946bf86d", |
| "test_multi_transformations.jsonl": "dcf88ab032056dc2b30605bc5cc50b9df81182bf7d390ca28f4ac8de61189b2e", |
| "test_multi_wording.jsonl": "a55114530eb38d5ec2337bb9899846a405ce0bc85ad5cb016d952256e2bd4e4f", |
| "test_original.jsonl": "b9767f8288b8a7f0eab9268f67105a0382b5e97a013d9e2523f9b144d5797368", |
| "test_single_branch.jsonl": "5dd117a8a60e22b337e911cec11138549565ca95db297ed57929a283d47c5a64", |
| "test_wording.jsonl": "9bee55f03bd059b59d32f430728d143d3fa41c16b60be7b76e444926901424ea", |
| "train_control.jsonl": "65fd1d4d8faebf404f5753483ee955251e4c5bbae91701bf5330fe7b410406f8", |
| "train_treatment.jsonl": "8af54c20ed280021a59a76d0b36690e08f12e5393aec0c67f638db78f476d8d5" |
| }, |
| "audit": { |
| "valid": true, |
| "families": 6996, |
| "questions": 45080, |
| "unique_worlds": 14366, |
| "prior_worlds_excluded": 100751, |
| "training_questions_per_arm": 12096, |
| "matched_label_stream": true, |
| "shared_training_questions": 6048, |
| "token_audit": { |
| "calibration": { |
| "questions": 1008, |
| "maximum_tokens": 271, |
| "total_input_tokens": 222101 |
| }, |
| "development": { |
| "questions": 1008, |
| "maximum_tokens": 271, |
| "total_input_tokens": 219347 |
| }, |
| "selection": { |
| "questions": 1008, |
| "maximum_tokens": 278, |
| "total_input_tokens": 219131 |
| }, |
| "selection_fit": { |
| "questions": 1008, |
| "maximum_tokens": 275, |
| "total_input_tokens": 222651 |
| }, |
| "test_counterfactual": { |
| "questions": 900, |
| "maximum_tokens": 278, |
| "total_input_tokens": 189531 |
| }, |
| "test_depth": { |
| "questions": 300, |
| "maximum_tokens": 333, |
| "total_input_tokens": 82274 |
| }, |
| "test_multi": { |
| "questions": 1440, |
| "maximum_tokens": 271, |
| "total_input_tokens": 332804 |
| }, |
| "test_multi_depth": { |
| "questions": 432, |
| "maximum_tokens": 353, |
| "total_input_tokens": 134163 |
| }, |
| "test_multi_signed_chain": { |
| "questions": 432, |
| "maximum_tokens": 257, |
| "total_input_tokens": 95379 |
| }, |
| "test_multi_transformations": { |
| "questions": 11520, |
| "maximum_tokens": 303, |
| "total_input_tokens": 2696668 |
| }, |
| "test_multi_wording": { |
| "questions": 432, |
| "maximum_tokens": 267, |
| "total_input_tokens": 98643 |
| }, |
| "test_original": { |
| "questions": 900, |
| "maximum_tokens": 278, |
| "total_input_tokens": 189486 |
| }, |
| "test_single_branch": { |
| "questions": 200, |
| "maximum_tokens": 244, |
| "total_input_tokens": 36305 |
| }, |
| "test_wording": { |
| "questions": 300, |
| "maximum_tokens": 261, |
| "total_input_tokens": 62899 |
| }, |
| "train_control": { |
| "questions": 12096, |
| "maximum_tokens": 280, |
| "total_input_tokens": 2535330 |
| }, |
| "train_treatment": { |
| "questions": 12096, |
| "maximum_tokens": 280, |
| "total_input_tokens": 2655786 |
| } |
| }, |
| "exclusion_files": { |
| "data/pilot-v3/calibration.jsonl": "f4fe9513ce84d49ca09f2bd3abb763c339bc87c92539a28c182538e1a758dddf", |
| "data/pilot-v3/development.jsonl": "89822945458c7652d9daa143224901ade591a6b9b3724b74bfd583b817e3cdea", |
| "data/pilot-v3/test_depth.jsonl": "cf825f3f48b50c868f0b400795648e724a239356f1b25135309b19ab00b37718", |
| "data/pilot-v3/test_id.jsonl": "186adf70fef3ec8a42cbaab221c374838979c936fcecc43340d3c7460b6aba87", |
| "data/pilot-v3/test_structure.jsonl": "e242154c0c0e57120c6f529748859d4a6f1239b98f53f772f13476950acda808", |
| "data/pilot-v3/test_wording.jsonl": "0d49676a5cbcf4963fc50b4dfb94c79e89a4bcb815c742d47c83bf0e9f057f0d", |
| "data/pilot-v3/train.jsonl": "c13ab0bea98f594aac6d99553b0cdfbe6840556e4bada5bfbcfdbfe605e93eed", |
| "runs/study-v1/test-inputs/flip_fact.jsonl": "39a111e0f7015b136e27dfde6cabd1fdf552da13c7e1aae530c0bb30afaff6d6", |
| "runs/study-v1/test-inputs/irrelevant.jsonl": "55e6a60998b45824fac24cb0c9c492681abe4d942d98e99ae0a393f3b116f219", |
| "runs/study-v1/test-inputs/paraphrase.jsonl": "2154fb016240bcf20e180ac326f3bbf7159eaba23b46487931974fce1c2b7b34", |
| "runs/study-v1/test-inputs/remove_premise.jsonl": "8037087ba59e2c8b713b3bff36a51d3dfe8d139d8ad2a76d11814d034df7cf44", |
| "runs/study-v1/test-inputs/rename.jsonl": "2d3667dabc9a8b16d1b7ff79ff2adffb66fcee57d93a5029d86cb4bdb25668a4", |
| "runs/study-v1/test-inputs/reorder.jsonl": "0fff378ade8b86d534a6103009583bb8313e3caa997dc61a4392f63867f42f0e", |
| "data/v2/control/calibration.jsonl": "5ad9e1723e2070062e82765b94f136fe53b7bae91529a5d6e416a7c13393dcf2", |
| "data/v2/control/development.jsonl": "8026630042839f44202aace250199fddd028f0dd45c918a0abe52b5d80f31c5c", |
| "data/v2/control/families.jsonl": "1c86773e9e19c4249ef346c62b5d6d6bfea64ff1a7a4e3df61aba65bd8b5c52f", |
| "data/v2/control/test_counterfactual.jsonl": "ca2d60dbc0966fec12a220396d3ce875bfd4aeec1b2563b2ebc82372774deae8", |
| "data/v2/control/test_depth.jsonl": "5d31d7c04f19000181fa826852e3f047b15adc022741ec53f0bbd878b18ceb9f", |
| "data/v2/control/test_original.jsonl": "dc4170ac42d7a6d38a86b050a1ff5f7321db5069077449ad89c3a3f22528833c", |
| "data/v2/control/test_single_branch.jsonl": "8ed6d94e46a89cb80e40f7deea315dbc8c6a485abff5803e1ac309481da65f17", |
| "data/v2/control/test_wording.jsonl": "0fca87f3888ac1e814736668ea4bb8522aca882b554311498468661b7f8818ab", |
| "data/v2/control/train.jsonl": "0f31d55ae347693dd21745ae07893dadb1e286fc888743ac9b25c5014eacb608", |
| "data/v2/treatment/calibration.jsonl": "5ad9e1723e2070062e82765b94f136fe53b7bae91529a5d6e416a7c13393dcf2", |
| "data/v2/treatment/development.jsonl": "8026630042839f44202aace250199fddd028f0dd45c918a0abe52b5d80f31c5c", |
| "data/v2/treatment/families.jsonl": "1c86773e9e19c4249ef346c62b5d6d6bfea64ff1a7a4e3df61aba65bd8b5c52f", |
| "data/v2/treatment/test_counterfactual.jsonl": "ca2d60dbc0966fec12a220396d3ce875bfd4aeec1b2563b2ebc82372774deae8", |
| "data/v2/treatment/test_depth.jsonl": "5d31d7c04f19000181fa826852e3f047b15adc022741ec53f0bbd878b18ceb9f", |
| "data/v2/treatment/test_original.jsonl": "dc4170ac42d7a6d38a86b050a1ff5f7321db5069077449ad89c3a3f22528833c", |
| "data/v2/treatment/test_single_branch.jsonl": "8ed6d94e46a89cb80e40f7deea315dbc8c6a485abff5803e1ac309481da65f17", |
| "data/v2/treatment/test_wording.jsonl": "0fca87f3888ac1e814736668ea4bb8522aca882b554311498468661b7f8818ab", |
| "data/v2/treatment/train.jsonl": "c008a0a7906ba56be85230907fc9de1d4177f4a9c8256bf2beb093a6fb5426c8", |
| "data/v2/families.jsonl": "1c86773e9e19c4249ef346c62b5d6d6bfea64ff1a7a4e3df61aba65bd8b5c52f", |
| "runs/study-v3/data/families.jsonl": "d939fbd5c98a4326547d8fead7b0a56ec3d6f208ccb1633602557d610904ae69", |
| "runs/study-v4/diagnostic/families.jsonl": "3904ebe4d5c2d8161f6276dc02466348ce09a6f19fdcf8531b6cad64adf333ed", |
| "runs/study-v4/training/families.jsonl": "fa19afb1379e632ef795ebc593299d9b590f9a6e90126bb5508154b9ce9a5559", |
| "runs/study-v5/data/families.jsonl": "27fa3adbdbb5192fac31263755d761ada825a5f53ed4d24c11fe956648c2bc1a", |
| "runs/study-v6/data/families.jsonl": "c45ee11a650b1e274c9685420f2551e0c155df98ba16aacb1d60948c4595c091" |
| }, |
| "preselected_family": "v7-r1-test_multi-multi-3-0000-0" |
| }, |
| "config_sha256": "29713ae51697f089b2aca6f7fec552e1f80109700b8f0fb2ba3a0b520482a388", |
| "lock_sha256": "7e57bf4c4a040908032b777dfbd1c18d2c567524528b512e987967a364792a76", |
| "protocol_markdown_sha256": "1e3c5a36e0056871fd32fc7fa46a04d7bccff28da0c1183d32604eb27159ea76", |
| "primary": "Treatment-minus-control complete-world accuracy on test_multi", |
| "selector": "Minimum calibrated selection NLL; earlier-step exact ties", |
| "test_predictions_inspected": false, |
| "identity": "f07e4f34618900fe4a571f47f380737efe7919897a5465704c92a138c3154c71" |
| } |
|
|
