AAMAS2027-Roles / VERSION
UHXinyang's picture ChihHsuan-Yang's picture
AAMAS2027-Roles v1.2.0
4790c54
Raw History Blame Contribute Delete
10.6 kB
{
"arm_a_differs_by_domain": "Arm A was DOMAIN-ADAPTED for SciBench (5 of 16 prompts differ, including the evaluator). Arm B is BYTE-IDENTICAL across domains. Do not assume one Arm A contract across the release, and do not compute P1a or any clause-specific fidelity metric across domains.",
"artifact": "AAMAS2027-Roles",
"artifact_version": "1.2.0",
"canonical_analysis": {
"note": "FINAL_TABLES.py describes the frozen campaign, not every result in the current release manuscript. Feedback-enabled controls are separate under controls/.",
"output": "analysis/out/FINAL_TABLES_OUTPUT.txt",
"script": "analysis/scripts/FINAL_TABLES.py",
"validation": "scripts/validate_artifact.py -- 181/181 PASS, re-run against the canonical output regenerated 2026-09-30 14:49Z"
},
"emitted_utc": "2026-10-31T23:05:00+00:00",
"frozen_at": "2026-09-30T10:09:30.201557",
"frozen_at_scibench": "2026-09-30T23:26:47.198588+00:00",
"licensing": "per-source and MIXED; see LICENSES/README.md and LICENSES/registry.json",
"mixed_licence": "Since v1.1.0 the release ships TWO upstreams under TWO licences: omnimath2 = Apache-2.0, scibench = MIT. Neither is applied over the other and the project's own Apache-2.0 covers neither. Filter by benchmark_id. LICENSES/ is the authority; the HF `license:` field can hold only one value.",
"release_history": [
{
"note": "first release; single domain (competition math, Omni-MATH-2). These three digests are UNCHANGED in 1.1.0 -- anything published against 1.0.0 remains checkable.",
"status": "private",
"traces": {
"actor3.jsonl": "fa22a446d5314bb67b1c6a6444b511eeeae3eb8770419a36f182617eeccbf3b5",
"armCp.jsonl": "038e08aec4d4aee166106a6325309dfe2301faf339c7d0bf82338a74576424d4",
"campaign_v2.jsonl": "e4e53cb117d91129991ce2fe87e336e152fc9ee6ca826d363f2bd5dd240d7130"
},
"version": "1.0.0"
},
{
"note": "ADDITIVE second domain (college chemistry/physics, SciBench): 2,296 trajectories, 574 problems, 2 actors, Arms A and B only. No v1.0.0 file changed bytes -- asserted by the build and by this emitter. SciBench is MIT, NOT the Apache-2.0 of the Omni-MATH material; see LICENSES/. ARM A's PROMPT DIFFERS BETWEEN DOMAINS (domain-adapted); Arm B is byte-identical.",
"status": "private",
"traces": {
"scibench_AB.jsonl": "10068f72cfd936d3be6de85fef196d63fc154ef2c2ec26a0ac2c7313513dce58"
},
"version": "1.1.0"
},
{
"version": "1.2.0",
"status": "public",
"date": "2026-10-31",
"note": "Redaction and completion release. (1) In 18 trajectory rows the `error` field held a response page returned by the inference gateway; each such value is replaced by a fixed marker (14 rows in raw/: campaign_v2 10, actor3 3, armCp 1; 4 rows in controls/run1/bf_omnimath_llama33.jsonl). Every other row is byte-identical. The original digests are kept as redacted_from_sha256; the original bytes are no longer distributed. (2) Adds controls/run1/: the first role-free-with-feedback (Bf) run in five cells, the gemma-4 arm-A re-run and the forced-submission run, which the paper's main text reports. (3) The parquet layer and raw/scibench_AB.jsonl are unchanged. (4) Internal release tooling removed (pre-release scanners, metadata emitters, release procedure).",
"traces": {
"actor3.jsonl": "687a389589a7d21fbc03bd65b22ad4d8a38ff5f29e5b902f5129f6e90b9d6435",
"armCp.jsonl": "63f8061c0db44ddae92881f5f5d06a60931e2383838659d7d98da89b0f90e2c6",
"campaign_v2.jsonl": "fb2fabf56a5e9066e75d916595e120a81b1343e48fc888da10e582460c8019b7"
},
"controls_added": [
"controls/run1/afs_omnimath_gemma4_n400.jsonl",
"controls/run1/bf_omnimath_gemma4.jsonl",
"controls/run1/bf_omnimath_gptoss.jsonl",
"controls/run1/bf_omnimath_llama33.jsonl",
"controls/run1/bf_scibench_gemma4.jsonl",
"controls/run1/bf_scibench_gptoss.jsonl",
"controls/run1/bridgeA_omnimath_gemma4_n400.jsonl"
]
}
],
"release_status": "PUBLIC_NAMED_RELEASE",
"released_traces": {
"actor3.jsonl": {
"actor_models": {
"meta-llama/Llama-3.3-70B-Instruct": 1200
},
"arms": {
"A": 400,
"B": 400,
"C": 400
},
"backend": [
"sophia_vllm"
],
"benchmark_id": "omnimath2",
"bytes": 46036452,
"config_version": {
"231b8c4fe205d04b": 1200
},
"domain": [
"math"
],
"first_started_at": "2026-09-29T15:47:04.182156+00:00",
"introduced_in_version": "1.0.0",
"last_started_at": "2026-09-29T20:44:16.485658+00:00",
"rows": 1200,
"run_id": [
"actor3-n400"
],
"semantic_version": "1.2.0",
"sha256": "687a389589a7d21fbc03bd65b22ad4d8a38ff5f29e5b902f5129f6e90b9d6435",
"redacted_from_sha256": "fa22a446d5314bb67b1c6a6444b511eeeae3eb8770419a36f182617eeccbf3b5",
"redacted_error_fields": 3
},
"armCp.jsonl": {
"actor_models": {
"google/gemma-4-31B-it": 400,
"openai/gpt-oss-120b": 400
},
"arms": {
"Cp": 800
},
"backend": [
"sophia_vllm"
],
"benchmark_id": "omnimath2",
"bytes": 23627512,
"config_version": {
"231b8c4fe205d04b": 800
},
"domain": [
"math"
],
"first_started_at": "2026-09-28T23:57:50.201018+00:00",
"introduced_in_version": "1.0.0",
"last_started_at": "2026-09-29T05:23:19.490039+00:00",
"rows": 800,
"run_id": [
"armCp-n400"
],
"semantic_version": "1.2.0",
"sha256": "63f8061c0db44ddae92881f5f5d06a60931e2383838659d7d98da89b0f90e2c6",
"redacted_from_sha256": "038e08aec4d4aee166106a6325309dfe2301faf339c7d0bf82338a74576424d4",
"redacted_error_fields": 1
},
"campaign_v2.jsonl": {
"actor_models": {
"google/gemma-4-31B-it": 4800,
"openai/gpt-oss-120b": 4800
},
"arms": {
"A": 3200,
"B": 3200,
"C": 3200
},
"backend": [
"sophia_vllm"
],
"benchmark_id": "omnimath2",
"bytes": 227523997,
"config_version": {
"e3afa39849c1b980": 9600
},
"domain": [
"math"
],
"first_started_at": "2026-09-28T23:26:56.570114+00:00",
"introduced_in_version": "1.0.0",
"last_started_at": "2026-09-29T22:29:37.082127+00:00",
"rows": 9600,
"run_id": [
"main-n1600-v2"
],
"semantic_version": "1.2.0",
"sha256": "fb2fabf56a5e9066e75d916595e120a81b1343e48fc888da10e582460c8019b7",
"redacted_from_sha256": "e4e53cb117d91129991ce2fe87e336e152fc9ee6ca826d363f2bd5dd240d7130",
"redacted_error_fields": 10
},
"scibench_AB.jsonl": {
"actor_models": {
"google/gemma-4-31B-it": 1148,
"openai/gpt-oss-120b": 1148
},
"arms": {
"A": 1148,
"B": 1148
},
"backend": [
"sophia_vllm"
],
"benchmark_id": "scibench",
"bytes": 26961648,
"config_version": {
"aed89f28ab338bdb": 2296
},
"domain": [
"science"
],
"first_started_at": "2026-09-30T21:38:50.509184+00:00",
"introduced_in_version": "1.1.0",
"last_started_at": "2026-09-30T23:21:41.275435+00:00",
"rows": 2296,
"run_id": [
"xdom-scibench-AB-full"
],
"semantic_version": "1.1.0",
"sha256": "10068f72cfd936d3be6de85fef196d63fc154ef2c2ec26a0ac2c7313513dce58"
}
},
"public_release_date": "2026-10-31",
"public_release_note": "Public release. Trajectory files are byte-identical to the frozen snapshot and the post-freeze runs except the 18 redacted `error` values listed under release_history 1.2.0. This is not a statement of conference acceptance.",
"control_traces": {
"controls/rolefree_feedback/gemma4/endpoint_drift_bridge/bridgeA_omnimath_gemma4_n400.jsonl": {
"sha256": "0ec97d88048e84e3d83f926bb6b42805d19061d359b0fa578fa334f3b4314f29",
"bytes": 8472059,
"rows": 400,
"manifest": "controls/rolefree_feedback/gemma4/MANIFEST.json"
},
"controls/rolefree_feedback/gemma4/omnimath/bf_omnimath_gemma4.jsonl": {
"sha256": "be3469f720048ca9aa21816ca6c9dd1f514170e3e38cffa2613db4317b1bb8e1",
"bytes": 27097121,
"rows": 1600,
"manifest": "controls/rolefree_feedback/gemma4/MANIFEST.json"
},
"controls/rolefree_feedback/gemma4/scibench/bf_scibench_gemma4.jsonl": {
"sha256": "965ed061ebb107d52a9063707deac7b43c3fcdffb698bf045af2b75e6d904b29",
"bytes": 8495787,
"rows": 574,
"manifest": "controls/rolefree_feedback/gemma4/MANIFEST.json"
},
"controls/run1/afs_omnimath_gemma4_n400.jsonl": {
"sha256": "47bed257b8783f9dfae177835b50dfdba19bab920a46d87ebd612f6cc60869cd",
"bytes": 6655595,
"rows": 400,
"manifest": "controls/run1/MANIFEST.json"
},
"controls/run1/bf_omnimath_gemma4.jsonl": {
"sha256": "d00cc8e3d09f50a2611e3ad211808d9f5cfa02737000f033cb0c4274511a3474",
"bytes": 27891526,
"rows": 1600,
"manifest": "controls/run1/MANIFEST.json"
},
"controls/run1/bf_omnimath_gptoss.jsonl": {
"sha256": "3dbfc93ab919d7c320f00f3c772568e5225d681d49e2863b786008b72fbd828e",
"bytes": 23796445,
"rows": 1600,
"manifest": "controls/run1/MANIFEST.json"
},
"controls/run1/bf_omnimath_llama33.jsonl": {
"sha256": "b53e730263db637bf0f3b564770e92148c430150f51965f19da57d39101022fb",
"bytes": 20288673,
"rows": 400,
"manifest": "controls/run1/MANIFEST.json",
"redacted_from_sha256": "42425019a7b44b9acb58dccb8788772e27047c5dbdc625a67f92880aad56de3d",
"redacted_error_fields": 4
},
"controls/run1/bf_scibench_gemma4.jsonl": {
"sha256": "494081dac0ecddc0875fc5440e331ec43654115a6afbbbd40829104047fa7224",
"bytes": 8430177,
"rows": 574,
"manifest": "controls/run1/MANIFEST.json"
},
"controls/run1/bf_scibench_gptoss.jsonl": {
"sha256": "e80d5c0928cead1935987d0342507e03c431f2cb3e482b0237a2ae865dab3cdd",
"bytes": 3553135,
"rows": 574,
"manifest": "controls/run1/MANIFEST.json"
},
"controls/run1/bridgeA_omnimath_gemma4_n400.jsonl": {
"sha256": "cfe9c8f80daca372ec640e83d887888e12b9e51196ed62325d77ea9d4940c717",
"bytes": 8820253,
"rows": 400,
"manifest": "controls/run1/MANIFEST.json"
}
}
}