Datasets:
traj_id stringlengths 19 19 | seed int64 20.4M 20.4M | split stringclasses 1
value | lang stringclasses 3
values | task_type stringclasses 5
values | goal stringlengths 82 195 | roles listlengths 2 4 | turns listlengths 6 16 | artifacts listlengths 2 5 | outcome dict | meta dict |
|---|---|---|---|---|---|---|---|---|---|---|
collab-train-000001 | 20,420,001 | train | mix | doc_align | 成功条件はREADME主張6項目をソースログと突合し、不一致を全件直して数値一致させる（対象: docs/readme_claims.csv、draft 666ac5） | [
{
"agent_id": "planner",
"name": "Minato",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Hina",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Tsumugi",
"brief": "誤り・リークを指摘。編集とツール不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "段取りはこうします。README主張6項目をソースログと突合し、不一致を全件直して数値一致させる。担当は coder、critic。よろしくお願いします。",
"tool": null,
"artifact_refs": [
"a1"
],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "作業結果の初... | [
{
"id": "a1",
"kind": "log",
"path": "logs/run_log.json",
"content_hash": "sha256:6a0b826aea5658d7f8dbdca31e03f77864d6295f1aed33d6d98b6f5c7e1507b1",
"summary": "ソースログ（toy）",
"content": "{\"dup_rate\": 0.25, \"frac_open\": 0.255, \"keep_rate\": 0.565, \"method\": \"soft-gate\", \"toy_mse_hard... | {
"success": true,
"metric": {
"name": "readme_mismatches",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 10,
"subtype": "method_name_mismatch",
"bank_id": "bank_train_v1",
"template_family_id": "doc_align/method_name_mismatch",
"attempt": 0,
"designed_negative": false,... |
collab-train-000002 | 20,420,002 | train | ja | bugfix | 成功条件は`chunk_sums` の不具合（空入力でZeroDivisionError）を1ファイル修正で直し、6ケースの実行テストを全通過させる（対象: src/chunk_sums.py、draft 9cd47a） | [
{
"agent_id": "planner",
"name": "Aoi",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Mei",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "verifier",
"name": "Tsumugi",
"brief": "検証器の実行のみ。編集不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "全体像から入ります。`chunk_sums` の不具合（空入力でZeroDivisionError）を1ファイル修正で直し、6ケースの実行テストを全通過させる。担当は coder、verifier。問題なければ進めます。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"ut... | [
{
"id": "a1",
"kind": "code",
"path": "src/chunk_sums.py",
"content_hash": "sha256:9cd47a50e1b8099fc2bf80d99a1dcc5824799328142748e718a3a59791c02fa6",
"summary": "初版",
"content": "def chunk_sums(x, k):\n return sum(x[-k:]) / len(x[-k:])\n"
},
{
"id": "a2",
"kind": "code",
"... | {
"success": true,
"metric": {
"name": "tests_passed_frac",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 8,
"subtype": "empty_input",
"bank_id": "bank_train_v1",
"template_family_id": "bugfix/empty_input",
"attempt": 0,
"designed_negative": false,
"final_artifact": "... |
collab-train-000003 | 20,420,003 | train | ja | synth_ablation | 成功条件は`accept_floor` の1因子表（4点・toy）を作り、c1<c2<…かつ cap=0.608 以下を数値assertで通す（対象: ablations/accept_floor.csv、draft d3371f） | [
{
"agent_id": "planner",
"name": "Tsumugi",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Hina",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Toa",
"brief": "誤り・リークを指摘。編集とツール不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "段取りはこうします。`accept_floor` の1因子表（4点・toy）を作り、c1<c2<…かつ cap=0.608 以下を数値assertで通す。担当は coder、critic。よろしくお願いします。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utteran... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/accept_floor.csv",
"content_hash": "sha256:d3371f3dc1d639c6530b6874aa842826e60201f36aeb62abe63e5ee11ba56d62",
"summary": "初版",
"content": "factor,value,toy_accept\naccept_floor,0.054,0.535\naccept_floor,0.258,-0.069\naccept_floor,0.463,0.291\... | {
"success": true,
"metric": {
"name": "strict_monotone_caps",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 8,
"subtype": "accept_positive",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/accept_positive",
"attempt": 0,
"designed_negative": false,
"fi... |
collab-train-000004 | 20,420,004 | train | ja | mqar_gen | 目標: MQAR縮小サンプル5件（keys=5, vocab=128）を生成し、集合演算検証の違反数を0にする（対象: mqar/samples.jsonl、draft 580e82） | [
{
"agent_id": "planner",
"name": "Minato",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Toa",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "verifier",
"name": "Hina",
"brief": "検証器の実行のみ。編集不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "段取りはこうします。MQAR縮小サンプル5件（keys=5, vocab=128）を生成し、集合演算検証の違反数を0にする。担当は coder、verifier。次に進めます。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "作業結果の初版です。5... | [
{
"id": "a1",
"kind": "code",
"path": "mqar/samples.jsonl",
"content_hash": "sha256:580e827ee97ff225210b0b9fcf5e1dc1068439216ceaa2e68fa2e41adce14e1e",
"summary": "初版",
"content": "{\"kv\": [[1, 119], [2, 116], [6, 114], [3, 96], [5, 118]], \"query\": [5, 3, 2, 1, 6], \"answer\": [118, 96, 11... | {
"success": true,
"metric": {
"name": "mqar_violations",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 10,
"subtype": "key_duplicate",
"bank_id": "bank_train_v1",
"template_family_id": "mqar_gen/key_duplicate",
"attempt": 0,
"designed_negative": false,
"final_artif... |
collab-train-000005 | 20,420,005 | train | ja | bugfix | 成功条件は`tail_mean` の不具合（最後の窓が落ちる（range終端が1つ短い））を1ファイル修正で直し、6ケースの実行テストを全通過させる（対象: src/tail_mean.py、draft 71e143） | [
{
"agent_id": "planner",
"name": "Rin",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Aoi",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Toa",
"brief": "誤り・リークを指摘。編集とツール不可"
},
{
"agent_id": "verifier",
"name": "M... | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "役割を割り振ります。`tail_mean` の不具合（最後の窓が落ちる（range終端が1つ短い））を1ファイル修正で直し、6ケースの実行テストを全通過させる。担当は coder、critic、verifier。次に進めます。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
... | [
{
"id": "a1",
"kind": "code",
"path": "src/tail_mean.py",
"content_hash": "sha256:71e1431c2478ef5d6e1d294f9cee33a73ac8a4f8f63b3f84cee87d1e8177d7bd",
"summary": "初版",
"content": "def tail_mean(x, k):\n return [sum(x[i:i+k]) for i in range(len(x)-k)]\n"
},
{
"id": "a2",
"kind": ... | {
"success": true,
"metric": {
"name": "tests_passed_frac",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 4,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 8,
"subtype": "off_by_one",
"bank_id": "bank_train_v1",
"template_family_id": "bugfix/off_by_one",
"attempt": 0,
"designed_negative": false,
"final_artifact": "a2... |
collab-train-000006 | 20,420,006 | train | en | mqar_gen | Done when we generate 3 mini-MQAR samples (keys=4, vocab=128) with zero violations under set-based checks (target: mqar/samples.jsonl, draft 1f5b80) | [
{
"agent_id": "planner",
"name": "Mei",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Yui",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Rin",
"brief": "誤り・リークを指摘。編集とツール不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "Let's confirm the steps. We generate 3 mini-MQAR samples (keys=4, vocab=128) with zero violations under set-based checks. Owners: coder, critic. See the log.",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
... | [
{
"id": "a1",
"kind": "code",
"path": "mqar/samples.jsonl",
"content_hash": "sha256:1f5b802b3657e1bdb4eb34329887302cc418c6cb643d2176c0500959c9db4246",
"summary": "first draft",
"content": "{\"kv\": [[0, 67], [5, 121], [4, 118], [1, 69]], \"query\": [5, 4], \"answer\": [121, 118]}\n{\"kv\": [... | {
"success": true,
"metric": {
"name": "mqar_violations",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 8,
"subtype": "key_duplicate",
"bank_id": "bank_train_v1",
"template_family_id": "mqar_gen/key_duplicate",
"attempt": 0,
"designed_negative": false,
"final_artifa... |
collab-train-000007 | 20,420,007 | train | ja | synth_ablation | 狙いは`accept_floor` の1因子表（7点・toy）を作り、c1<c2<…かつ cap=0.653 以下を数値assertで通す（対象: ablations/accept_floor.csv、draft 96ab0c） | [
{
"agent_id": "planner",
"name": "Tsumugi",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Mei",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Emi",
"brief": "誤り・リークを指摘。編集とツール不可"
},
{
"agent_id": "verifier",
"name"... | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "計画を置きます。`accept_floor` の1因子表（7点・toy）を作り、c1<c2<…かつ cap=0.653 以下を数値assertで通す。担当は coder、critic、verifier。続けます。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"uttera... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/accept_floor.csv",
"content_hash": "sha256:96ab0c6963cb47a4352e09af31db7e56e9ff236964ade635d844254a3d3407c5",
"summary": "初版",
"content": "factor,value,toy_accept\naccept_floor,0.107,0.187\naccept_floor,0.245,0.367\naccept_floor,0.382,0.424\n... | {
"success": true,
"metric": {
"name": "strict_monotone_caps",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 4,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 11,
"subtype": "accept_positive",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/accept_positive",
"attempt": 0,
"designed_negative": false,
"f... |
collab-train-000008 | 20,420,008 | train | ja | protocol_review | 到達点はプロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする（対象: protocols/eval_protocol.md、draft 47135a） | [
{
"agent_id": "planner",
"name": "Minato",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Hina",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "計画を置きます。プロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする。担当は coder。問題なければ進めます。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "手を動かしました。帯を CALIB 20880001-... | [
{
"id": "a1",
"kind": "protocol",
"path": "protocols/eval_protocol.md",
"content_hash": "sha256:47135a74029371a0372b166cf1c38d1c8bfdd8dd448e3bc2accc731b6e95f780",
"summary": "初版",
"content": "CALIB: 20880001-20880023 (thresholds only)\nTEST: 20880032-20880097\nHOLD: 20880105-20880146 used fo... | {
"success": false,
"metric": {
"name": "protocol_violations",
"value": 1
},
"stopped_reason": "budget"
} | {
"max_turns": 16,
"n_agents": 2,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 8,
"subtype": "holdout_go_unused",
"bank_id": "bank_train_v1",
"template_family_id": "protocol_review/holdout_go_unused",
"attempt": 2,
"designed_negative": true,
... |
collab-train-000009 | 20,420,009 | train | ja | doc_align | 到達点はREADME主張4項目をソースログと突合し、不一致を全件直して数値一致させる（対象: docs/readme_claims.csv、draft ebc9fa） | [
{
"agent_id": "planner",
"name": "Aoi",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Yui",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "verifier",
"name": "Rin",
"brief": "検証器の実行のみ。編集不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "役割を割り振ります。README主張4項目をソースログと突合し、不一致を全件直して数値一致させる。担当は coder、verifier。記録に残します。",
"tool": null,
"artifact_refs": [
"a1"
],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "ask",
"utterance": "一つ質問です。外部... | [
{
"id": "a1",
"kind": "log",
"path": "logs/run_log.json",
"content_hash": "sha256:56dec480fd3a2da5b50cd0f2163bc561c97cd4dc1b32d3c34283bd0b711bf5e9",
"summary": "ソースログ（toy）",
"content": "{\"dup_rate\": 0.164, \"method\": \"soft-gate\", \"toy_mse_hard\": 0.371, \"toy_mse_soft\": 0.521}"
},
... | {
"success": true,
"metric": {
"name": "readme_mismatches",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 16,
"subtype": "method_name_mismatch",
"bank_id": "bank_train_v1",
"template_family_id": "doc_align/method_name_mismatch",
"attempt": 0,
"designed_negative": false,... |
collab-train-000010 | 20,420,010 | train | mix | protocol_review | 狙いはプロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする（対象: protocols/eval_protocol.md、draft 2ebe06） | [
{
"agent_id": "planner",
"name": "Tsumugi",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Emi",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "計画を置きます。プロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする。担当は coder。異論あればどうぞ。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "叩き台です。帯を CALIB 20820001-20820... | [
{
"id": "a1",
"kind": "protocol",
"path": "protocols/eval_protocol.md",
"content_hash": "sha256:2ebe0685b408ce277bc47d48be43e26970209819597c11018b51f0f53041abc1",
"summary": "初版",
"content": "CALIB: 20820001-20820018 (thresholds only)\nTEST: 20820032-20820121\nHOLD: 20820126-20820178 (sealed... | {
"success": true,
"metric": {
"name": "protocol_violations",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 2,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 6,
"subtype": "test_threshold_tuning",
"bank_id": "bank_train_v1",
"template_family_id": "protocol_review/test_threshold_tuning",
"attempt": 0,
"designed_negative":... |
collab-train-000011 | 20,420,011 | train | ja | synth_ablation | 到達点は`rho` の1因子表（5点・toy）を作り、c1<c2<…かつ cap=0.64 以下を数値assertで通す（対象: ablations/rho.csv、draft 71108c） | [
{
"agent_id": "planner",
"name": "Hina",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Toa",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "verifier",
"name": "Yui",
"brief": "検証器の実行のみ。編集不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "計画を置きます。`rho` の1因子表（5点・toy）を作り、c1<c2<…かつ cap=0.64 以下を数値assertで通す。担当は coder、verifier。次に進めます。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "まず一案を置きま... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/rho.csv",
"content_hash": "sha256:71108c4111d8855e180211e0909fc3c70cd7cf2045b45db8f53b07bb2be1d9e4",
"summary": "初版",
"content": "factor,value,toy_accept\nrho,0.112,0.499\nrho,0.272,0.257\nrho,0.432,0.553\nrho,0.591,0.577\nrho,0.751,0.561"
... | {
"success": false,
"metric": {
"name": "strict_monotone_caps",
"value": 0
},
"stopped_reason": "budget"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 6,
"subtype": "rho_monotone",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/rho_monotone",
"attempt": 0,
"designed_negative": true,
"final_art... |
collab-train-000012 | 20,420,012 | train | mix | protocol_review | 成功条件はプロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする（対象: protocols/eval_protocol.md、draft d4c45f） | [
{
"agent_id": "planner",
"name": "Rin",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Hina",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Toa",
"brief": "誤り・リークを指摘。編集とツール不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "全体像から入ります。プロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする。担当は coder、critic。合意できれば次へ。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "作業結果の初版です。帯を CALIB 2... | [
{
"id": "a1",
"kind": "protocol",
"path": "protocols/eval_protocol.md",
"content_hash": "sha256:d4c45fca47d9eb7d73e4cbd5c6c16eb8f518ae8956be1843bcd48629901439a5",
"summary": "初版",
"content": "CALIB: 20680001-20680036 (thresholds only)\nTEST: 20680048-20680157\nHOLD: 20680172-20680216 used fo... | {
"success": true,
"metric": {
"name": "protocol_violations",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 14,
"subtype": "holdout_go_unused",
"bank_id": "bank_train_v1",
"template_family_id": "protocol_review/holdout_go_unused",
"attempt": 0,
"designed_negative": false,... |
collab-train-000013 | 20,420,013 | train | en | synth_ablation | Objective: build a 6-point toy table for `accept_floor` that passes numeric asserts c1<c2<… and ≤ cap=0.66 (target: ablations/accept_floor.csv, draft a5d0e4) | [
{
"agent_id": "planner",
"name": "Tsumugi",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Mei",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Aoi",
"brief": "誤り・リークを指摘。編集とツール不可"
},
{
"agent_id": "verifier",
"name"... | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "Splitting the work first. We build a 6-point toy table for `accept_floor` that passes numeric asserts c1<c2<… and ≤ cap=0.66. Owners: coder, critic, verifier. See the log.",
"tool": null,
"artifact_refs": [],
"claims": []
},
{... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/accept_floor.csv",
"content_hash": "sha256:a5d0e4d3a26706fd24c2d596d7915ec0e12e0702aefe273411f942cf664acb0e",
"summary": "first draft",
"content": "factor,value,toy_accept\naccept_floor,0.231,0.45\naccept_floor,0.375,0.36\naccept_floor,0.519,... | {
"success": true,
"metric": {
"name": "strict_monotone_caps",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 4,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 16,
"subtype": "accept_positive",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/accept_positive",
"attempt": 0,
"designed_negative": false,
"f... |
collab-train-000014 | 20,420,014 | train | ja | protocol_review | 成功条件はプロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする（対象: protocols/eval_protocol.md、draft 5486b4） | [
{
"agent_id": "planner",
"name": "Emi",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Toa",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "計画を置きます。プロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする。担当は coder。記録に残します。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "まず一案を置きます。帯を CALIB 20970001-20... | [
{
"id": "a1",
"kind": "protocol",
"path": "protocols/eval_protocol.md",
"content_hash": "sha256:5486b40f6530244ab43274e8b2ef382b7325927fc2b04627e9e47b9d5fb231f0",
"summary": "初版",
"content": "CALIB: 20970001-20970037 (thresholds only)\nTEST: 20970053-20970103\nHOLD: 20970105-20970160 (sealed... | {
"success": true,
"metric": {
"name": "protocol_violations",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 2,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 9,
"subtype": "test_threshold_tuning",
"bank_id": "bank_train_v1",
"template_family_id": "protocol_review/test_threshold_tuning",
"attempt": 0,
"designed_negative":... |
collab-train-000015 | 20,420,015 | train | ja | doc_align | 到達点はREADME主張5項目をソースログと突合し、不一致を全件直して数値一致させる（対象: docs/readme_claims.csv、draft 725dce） | [
{
"agent_id": "planner",
"name": "Hina",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Emi",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Aoi",
"brief": "誤り・リークを指摘。編集とツール不可"
},
{
"agent_id": "verifier",
"name": "... | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "役割を割り振ります。README主張5項目をソースログと突合し、不一致を全件直して数値一致させる。担当は coder、critic、verifier。合意できれば次へ。",
"tool": null,
"artifact_refs": [
"a1"
],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": ... | [
{
"id": "a1",
"kind": "log",
"path": "logs/run_log.json",
"content_hash": "sha256:645b677166d1cfec9f430b3ac193faa2a351d7f588e1d0057248b5873bf2caf2",
"summary": "ソースログ（toy）",
"content": "{\"archive_bytes\": 607113, \"dup_rate\": 0.069, \"frac_open\": 0.515, \"mean_g\": 0.284, \"toy_mse_hard\"... | {
"success": true,
"metric": {
"name": "readme_mismatches",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 4,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 12,
"subtype": "kib_calc",
"bank_id": "bank_train_v1",
"template_family_id": "doc_align/kib_calc",
"attempt": 0,
"designed_negative": false,
"final_artifact": "a5... |
collab-train-000016 | 20,420,016 | train | ja | mqar_gen | 目標: MQAR縮小サンプル3件（keys=5, vocab=256）を生成し、集合演算検証の違反数を0にする（対象: mqar/samples.jsonl、draft 8d00d3） | [
{
"agent_id": "planner",
"name": "Aoi",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Emi",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "計画を置きます。MQAR縮小サンプル3件（keys=5, vocab=256）を生成し、集合演算検証の違反数を0にする。担当は coder。合意できれば次へ。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "ask",
"utterance": "一つ質問です。TESTを見て閾値を動かしてよ... | [
{
"id": "a1",
"kind": "code",
"path": "mqar/samples.jsonl",
"content_hash": "sha256:8d00d3d74626aeb3a82fba35c6c38471e0583e50f09a04d716fe5424ea29327c",
"summary": "初版",
"content": "{\"kv\": [[83, 191], [46, 159], [85, 221], [36, 226], [123, 204]], \"query\": [124, 46, 123, 85, 36], \"answer\"... | {
"success": true,
"metric": {
"name": "mqar_violations",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 2,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 13,
"subtype": "pi_exact_cell",
"bank_id": "bank_train_v1",
"template_family_id": "mqar_gen/pi_exact_cell",
"attempt": 0,
"designed_negative": false,
"final_artif... |
collab-train-000017 | 20,420,017 | train | ja | protocol_review | 到達点はプロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする（対象: protocols/eval_protocol.md、draft 027833） | [
{
"agent_id": "planner",
"name": "Mei",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Toa",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Minato",
"brief": "誤り・リークを指摘。編集とツール不可"
},
{
"agent_id": "verifier",
"name":... | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "役割を割り振ります。プロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする。担当は coder、critic、verifier。根拠はartifactにあります。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "手を動... | [
{
"id": "a1",
"kind": "protocol",
"path": "protocols/eval_protocol.md",
"content_hash": "sha256:0278337d5adda215024fa737716e3b8c5ad721f628390e8db48c375da6a4f54a",
"summary": "初版",
"content": "CALIB: 20530001-20530023 (thresholds only)\nTEST: 20530022-20530076\nHOLD: 20530090-20530130 (sealed... | {
"success": true,
"metric": {
"name": "protocol_violations",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 4,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 9,
"subtype": "seed_overlap",
"bank_id": "bank_train_v1",
"template_family_id": "protocol_review/seed_overlap",
"attempt": 0,
"designed_negative": false,
"final_a... |
collab-train-000018 | 20,420,018 | train | ja | bugfix | 狙いは`window_sums` の不具合（空入力でZeroDivisionError）を1ファイル修正で直し、6ケースの実行テストを全通過させる（対象: src/window_sums.py、draft ab5613） | [
{
"agent_id": "planner",
"name": "Rin",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Mei",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "全体像から入ります。`window_sums` の不具合（空入力でZeroDivisionError）を1ファイル修正で直し、6ケースの実行テストを全通過させる。担当は coder。根拠はartifactにあります。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utte... | [
{
"id": "a1",
"kind": "code",
"path": "src/window_sums.py",
"content_hash": "sha256:ab5613efc2b4b800c6692e70c875701924d9289fc2bb9fe91a7e53a129618778",
"summary": "初版",
"content": "def window_sums(x, k):\n return sum(x[-k:]) / len(x[-k:])\n"
},
{
"id": "a2",
"kind": "code",
... | {
"success": true,
"metric": {
"name": "tests_passed_frac",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 2,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 6,
"subtype": "empty_input",
"bank_id": "bank_train_v1",
"template_family_id": "bugfix/empty_input",
"attempt": 0,
"designed_negative": false,
"final_artifact": "... |
collab-train-000019 | 20,420,019 | train | en | doc_align | Done when we cross-check 4 README claims against the source log and correct every mismatch until values match (target: docs/readme_claims.csv, draft f57383) | [
{
"agent_id": "planner",
"name": "Yui",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Minato",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "verifier",
"name": "Aoi",
"brief": "検証器の実行のみ。編集不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "Here's the sequence. We cross-check 4 README claims against the source log and correct every mismatch until values match. Owners: coder, verifier. See the log.",
"tool": null,
"artifact_refs": [
"a1"
],
"claims": []
},... | [
{
"id": "a1",
"kind": "log",
"path": "logs/run_log.json",
"content_hash": "sha256:d117ceaf7e261c515933074404b8fe2f70071d1413f97c30d582a1ccae365e1a",
"summary": "ソースログ（toy）",
"content": "{\"archive_bytes\": 863745, \"keep_rate\": 0.195, \"mean_g\": 0.373, \"toy_mse_hard\": 0.537}"
},
{
... | {
"success": true,
"metric": {
"name": "readme_mismatches",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 8,
"subtype": "kib_calc",
"bank_id": "bank_train_v1",
"template_family_id": "doc_align/kib_calc",
"attempt": 0,
"designed_negative": false,
"final_artifact": "a3"... |
collab-train-000020 | 20,420,020 | train | ja | bugfix | 目標: `tail_mean` の不具合（最後の窓が落ちる（range終端が1つ短い））を1ファイル修正で直し、6ケースの実行テストを全通過させる（対象: src/tail_mean.py、draft 71e143） | [
{
"agent_id": "planner",
"name": "Tsumugi",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Emi",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "verifier",
"name": "Toa",
"brief": "検証器の実行のみ。編集不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "役割を割り振ります。`tail_mean` の不具合（最後の窓が落ちる（range終端が1つ短い））を1ファイル修正で直し、6ケースの実行テストを全通過させる。担当は coder、verifier。記録に残します。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "ask",
"utteran... | [
{
"id": "a1",
"kind": "code",
"path": "src/tail_mean.py",
"content_hash": "sha256:71e1431c2478ef5d6e1d294f9cee33a73ac8a4f8f63b3f84cee87d1e8177d7bd",
"summary": "初版",
"content": "def tail_mean(x, k):\n return [sum(x[i:i+k]) for i in range(len(x)-k)]\n"
},
{
"id": "a2",
"kind": ... | {
"success": true,
"metric": {
"name": "tests_passed_frac",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 9,
"subtype": "off_by_one",
"bank_id": "bank_train_v1",
"template_family_id": "bugfix/off_by_one",
"attempt": 0,
"designed_negative": false,
"final_artifact": "a2... |
collab-train-000021 | 20,420,021 | train | en | synth_ablation | Done when we build a 5-point toy table for `tau` that passes numeric asserts c1<c2<… and ≤ cap=0.621 (target: ablations/tau.csv, draft 5b5426) | [
{
"agent_id": "planner",
"name": "Emi",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Rin",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Tsumugi",
"brief": "誤り・リークを指摘。編集とツール不可"
},
{
"agent_id": "verifier",
"name"... | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "Here's the sequence. We build a 5-point toy table for `tau` that passes numeric asserts c1<c2<… and ≤ cap=0.621. Owners: coder, critic, verifier. Should be self-explanatory.",
"tool": null,
"artifact_refs": [],
"claims": []
},
... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/tau.csv",
"content_hash": "sha256:5b54264dbda71480f3234c4dad5644e10ac82c29f4c8a82ad0cb0a42ea863f90",
"summary": "first draft",
"content": "factor,value,toy_accept\ntau,0.179,0.371\ntau,0.395,0.502\ntau,0.503,0.696\ntau,0.611,0.562"
},
{
... | {
"success": true,
"metric": {
"name": "strict_monotone_caps",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 4,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 7,
"subtype": "tau_grid",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/tau_grid",
"attempt": 0,
"designed_negative": false,
"final_artifact":... |
collab-train-000022 | 20,420,022 | train | ja | synth_ablation | 目標: `accept_floor` の1因子表（7点・toy）を作り、c1<c2<…かつ cap=0.637 以下を数値assertで通す（対象: ablations/accept_floor.csv、draft 3a0d13） | [
{
"agent_id": "planner",
"name": "Tsumugi",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Minato",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "段取りはこうします。`accept_floor` の1因子表（7点・toy）を作り、c1<c2<…かつ cap=0.637 以下を数値assertで通す。担当は coder。根拠はartifactにあります。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utteranc... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/accept_floor.csv",
"content_hash": "sha256:3a0d1357145e65a68e8f8b80b8ff557dc4a9aa48714f34ed1104daa25449a255",
"summary": "初版",
"content": "factor,value,toy_accept\naccept_floor,0.21,0.123\naccept_floor,0.315,0.481\naccept_floor,0.421,0.455\na... | {
"success": true,
"metric": {
"name": "strict_monotone_caps",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 2,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 10,
"subtype": "accept_positive",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/accept_positive",
"attempt": 0,
"designed_negative": false,
"f... |
collab-train-000023 | 20,420,023 | train | mix | mqar_gen | 目標: MQAR縮小サンプル3件（keys=5, vocab=64）を生成し、集合演算検証の違反数を0にする（対象: mqar/samples.jsonl、draft 8a2105） | [
{
"agent_id": "planner",
"name": "Hina",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Tsumugi",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "段取りはこうします。MQAR縮小サンプル3件（keys=5, vocab=64）を生成し、集合演算検証の違反数を0にする。担当は coder。根拠はartifactにあります。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "手を動かしました。3件... | [
{
"id": "a1",
"kind": "code",
"path": "mqar/samples.jsonl",
"content_hash": "sha256:8a2105c52263723d837f8545a52529f44638696c1671c437b2e246375d8d8854",
"summary": "初版",
"content": "{\"kv\": [[2, 46], [3, 57], [5, 35], [1, 59], [0, 56]], \"query\": [5, 2, 1, 0], \"answer\": [35, 46, 59, 56]}\n... | {
"success": true,
"metric": {
"name": "mqar_violations",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 2,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 10,
"subtype": "key_duplicate",
"bank_id": "bank_train_v1",
"template_family_id": "mqar_gen/key_duplicate",
"attempt": 0,
"designed_negative": false,
"final_artif... |
collab-train-000024 | 20,420,024 | train | ja | mqar_gen | 成功条件はMQAR縮小サンプル4件（keys=3, vocab=128）を生成し、集合演算検証の違反数を0にする（対象: mqar/samples.jsonl、draft e37a2d） | [
{
"agent_id": "planner",
"name": "Tsumugi",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Aoi",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Mei",
"brief": "誤り・リークを指摘。編集とツール不可"
},
{
"agent_id": "verifier",
"name"... | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "全体像から入ります。MQAR縮小サンプル4件（keys=3, vocab=128）を生成し、集合演算検証の違反数を0にする。担当は coder、critic、verifier。合意できれば次へ。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "ask",
"utterance": "質問があ... | [
{
"id": "a1",
"kind": "code",
"path": "mqar/samples.jsonl",
"content_hash": "sha256:e37a2d6086cf19a5991140eafc231f78e0b14d6950186168bb7cceaf2d49a231",
"summary": "初版",
"content": "{\"kv\": [[4, 8], [0, 12], [2, 15]], \"query\": [0, 2], \"answer\": [12, 15]}\n{\"kv\": [[0, 12], [1, 15], [4, 1... | {
"success": false,
"metric": {
"name": "mqar_violations",
"value": 11
},
"stopped_reason": "budget"
} | {
"max_turns": 16,
"n_agents": 4,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 14,
"subtype": "vocab_shortage",
"bank_id": "bank_train_v1",
"template_family_id": "mqar_gen/vocab_shortage",
"attempt": 0,
"designed_negative": true,
"final_arti... |
collab-train-000025 | 20,420,025 | train | en | synth_ablation | Done when we build a 6-point toy table for `rho` that passes numeric asserts c1<c2<… and ≤ cap=0.658 (target: ablations/rho.csv, draft 1bee08) | [
{
"agent_id": "planner",
"name": "Yui",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Minato",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Hina",
"brief": "誤り・リークを指摘。編集とツール不可"
},
{
"agent_id": "verifier",
"name"... | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "Let's confirm the steps. We build a 6-point toy table for `rho` that passes numeric asserts c1<c2<… and ≤ cap=0.658. Owners: coder, critic, verifier. See the log.",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t":... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/rho.csv",
"content_hash": "sha256:1bee080e2ff4f093fbb7197e49ef73862d35da977530700d9daaf5d002a2e973",
"summary": "first draft",
"content": "factor,value,toy_accept\nrho,0.057,0.3\nrho,0.178,0.192\nrho,0.3,0.392\nrho,0.421,0.346\nrho,0.542,0.50... | {
"success": true,
"metric": {
"name": "strict_monotone_caps",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 4,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 9,
"subtype": "rho_monotone",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/rho_monotone",
"attempt": 0,
"designed_negative": false,
"final_ar... |
collab-train-000026 | 20,420,026 | train | en | synth_ablation | Objective: build a 6-point toy table for `tau` that passes numeric asserts c1<c2<… and ≤ cap=0.828 (target: ablations/tau.csv, draft f764aa) | [
{
"agent_id": "planner",
"name": "Minato",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Emi",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Tsumugi",
"brief": "誤り・リークを指摘。編集とツール不可"
},
{
"agent_id": "verifier",
"na... | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "Starting from the big picture. We build a 6-point toy table for `tau` that passes numeric asserts c1<c2<… and ≤ cap=0.828. Owners: coder, critic, verifier. Done for now.",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/tau.csv",
"content_hash": "sha256:f764aaf9a1b3fb32deaad77c7338d401bb0224ea250de01e7419735d9bd7ac29",
"summary": "first draft",
"content": "factor,value,toy_accept\ntau,0.12,0.223\ntau,0.259,0.581\ntau,0.538,0.357\ntau,0.677,0.368\ntau,0.816,0... | {
"success": true,
"metric": {
"name": "strict_monotone_caps",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 4,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 7,
"subtype": "tau_grid",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/tau_grid",
"attempt": 0,
"designed_negative": false,
"final_artifact":... |
collab-train-000027 | 20,420,027 | train | ja | synth_ablation | 到達点は`accept_floor` の1因子表（5点・toy）を作り、c1<c2<…かつ cap=0.854 以下を数値assertで通す（対象: ablations/accept_floor.csv、draft 02c235） | [
{
"agent_id": "planner",
"name": "Hina",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Toa",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Tsumugi",
"brief": "誤り・リークを指摘。編集とツール不可"
},
{
"agent_id": "verifier",
"name... | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "役割を割り振ります。`accept_floor` の1因子表（5点・toy）を作り、c1<c2<…かつ cap=0.854 以下を数値assertで通す。担当は coder、critic、verifier。よろしくお願いします。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "ask",
"... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/accept_floor.csv",
"content_hash": "sha256:02c23504e2224cb9f020ac641c351c04f171749d0e67a196bfb1e07dff4c3c0c",
"summary": "初版",
"content": "factor,value,toy_accept\naccept_floor,0.236,0.189\naccept_floor,0.341,0.547\naccept_floor,0.447,-0.08\n... | {
"success": true,
"metric": {
"name": "strict_monotone_caps",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 4,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 13,
"subtype": "accept_positive",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/accept_positive",
"attempt": 0,
"designed_negative": false,
"f... |
collab-train-000028 | 20,420,028 | train | ja | protocol_review | 到達点はプロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする（対象: protocols/eval_protocol.md、draft f9230d） | [
{
"agent_id": "planner",
"name": "Minato",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Rin",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "計画を置きます。プロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする。担当は coder。合意できれば次へ。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "まず一案を置きます。帯を CALIB 20850001-2... | [
{
"id": "a1",
"kind": "protocol",
"path": "protocols/eval_protocol.md",
"content_hash": "sha256:f9230dd9c5204a94de2d4c3b86fedbf522f6a38dbd8a8ed86b6f68bc974a479e",
"summary": "初版",
"content": "CALIB: 20850001-20850019 (thresholds only)\nTEST: 20850031-20850134\nHOLD: 20850139-20850185 used fo... | {
"success": true,
"metric": {
"name": "protocol_violations",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 2,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 10,
"subtype": "holdout_go_unused",
"bank_id": "bank_train_v1",
"template_family_id": "protocol_review/holdout_go_unused",
"attempt": 0,
"designed_negative": false,... |
collab-train-000029 | 20,420,029 | train | ja | synth_ablation | 目標: `rho` の1因子表（5点・toy）を作り、c1<c2<…かつ cap=0.912 以下を数値assertで通す（対象: ablations/rho.csv、draft 395f34） | [
{
"agent_id": "planner",
"name": "Rin",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Tsumugi",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "verifier",
"name": "Aoi",
"brief": "検証器の実行のみ。編集不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "段取りはこうします。`rho` の1因子表（5点・toy）を作り、c1<c2<…かつ cap=0.912 以下を数値assertで通す。担当は coder、verifier。異論あればどうぞ。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "ask",
"utterance": "質問があり... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/rho.csv",
"content_hash": "sha256:395f345d65a8ed7f3f639d2dbba7e419009dc34058c3089312dba013f5596572",
"summary": "初版",
"content": "factor,value,toy_accept\nrho,0.156,0.6030000000000001\nrho,0.341,0.142\nrho,0.525,0.478\nrho,0.709,0.654\nrho,0.... | {
"success": true,
"metric": {
"name": "strict_monotone_caps",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 12,
"subtype": "rho_monotone",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/rho_monotone",
"attempt": 0,
"designed_negative": false,
"final_a... |
collab-train-000030 | 20,420,030 | train | ja | protocol_review | 狙いはプロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする（対象: protocols/eval_protocol.md、draft eeffb7） | [
{
"agent_id": "planner",
"name": "Mei",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Minato",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "段取りはこうします。プロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする。担当は coder。異論あればどうぞ。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "作業結果の初版です。帯を CALIB 20650001... | [
{
"id": "a1",
"kind": "protocol",
"path": "protocols/eval_protocol.md",
"content_hash": "sha256:eeffb75315338b8dcdb3c1b968a6b89677ab8eb3223056f0a7fad32e3c7c4239",
"summary": "初版",
"content": "CALIB: 20650001-20650037 (thresholds only)\nTEST: 20650040-20650114\nHOLD: 20650131-20650174 used fo... | {
"success": false,
"metric": {
"name": "protocol_violations",
"value": 1
},
"stopped_reason": "budget"
} | {
"max_turns": 16,
"n_agents": 2,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 8,
"subtype": "holdout_go_unused",
"bank_id": "bank_train_v1",
"template_family_id": "protocol_review/holdout_go_unused",
"attempt": 0,
"designed_negative": true,
... |
collab-train-000031 | 20,420,031 | train | ja | synth_ablation | 到達点は`accept_floor` の1因子表（7点・toy）を作り、c1<c2<…かつ cap=0.769 以下を数値assertで通す（対象: ablations/accept_floor.csv、draft 7b7370） | [
{
"agent_id": "planner",
"name": "Minato",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Hina",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "役割を割り振ります。`accept_floor` の1因子表（7点・toy）を作り、c1<c2<…かつ cap=0.769 以下を数値assertで通す。担当は coder。異論あればどうぞ。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "まず一... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/accept_floor.csv",
"content_hash": "sha256:7b73702d91412e3c60756b5323421363afee57fc600a4f60703e1dbef5b58469",
"summary": "初版",
"content": "factor,value,toy_accept\naccept_floor,0.211,-0.025\naccept_floor,0.333,0.437\naccept_floor,0.455,0.503\... | {
"success": true,
"metric": {
"name": "strict_monotone_caps",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 2,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 9,
"subtype": "accept_positive",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/accept_positive",
"attempt": 0,
"designed_negative": false,
"fi... |
collab-train-000032 | 20,420,032 | train | ja | synth_ablation | 狙いは`accept_floor` の1因子表（4点・toy）を作り、c1<c2<…かつ cap=0.791 以下を数値assertで通す（対象: ablations/accept_floor.csv、draft f2f82e） | [
{
"agent_id": "planner",
"name": "Minato",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Emi",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Yui",
"brief": "誤り・リークを指摘。編集とツール不可"
},
{
"agent_id": "verifier",
"name":... | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "全体像から入ります。`accept_floor` の1因子表（4点・toy）を作り、c1<c2<…かつ cap=0.791 以下を数値assertで通す。担当は coder、critic、verifier。次に進めます。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"ut... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/accept_floor.csv",
"content_hash": "sha256:f2f82e092f519becb2a617587947b7be17f173b2cddf7f4d5c3d68f665527e84",
"summary": "初版",
"content": "factor,value,toy_accept\naccept_floor,0.166,0.14\naccept_floor,0.403,0.574\naccept_floor,0.64,-0.022\na... | {
"success": true,
"metric": {
"name": "strict_monotone_caps",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 4,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 11,
"subtype": "accept_positive",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/accept_positive",
"attempt": 0,
"designed_negative": false,
"f... |
collab-train-000033 | 20,420,033 | train | en | synth_ablation | Done when we build a 4-point toy table for `tau` that passes numeric asserts c1<c2<… and ≤ cap=0.604 (target: ablations/tau.csv, draft aea539) | [
{
"agent_id": "planner",
"name": "Minato",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Yui",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Toa",
"brief": "誤り・リークを指摘。編集とツール不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "Here's the sequence. We build a 4-point toy table for `tau` that passes numeric asserts c1<c2<… and ≤ cap=0.604. Owners: coder, critic. Short, but that's all.",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/tau.csv",
"content_hash": "sha256:aea5391606ef0fdc61fa7684dafd3af502343cdd2ff270b8d3d9f39e3da1f7ab",
"summary": "first draft",
"content": "factor,value,toy_accept\ntau,0.214,0.347\ntau,0.413,0.504\ntau,0.81,0.445"
},
{
"id": "a2",
... | {
"success": true,
"metric": {
"name": "strict_monotone_caps",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 7,
"subtype": "tau_grid",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/tau_grid",
"attempt": 0,
"designed_negative": false,
"final_artifact":... |
collab-train-000034 | 20,420,034 | train | ja | doc_align | 成功条件はREADME主張5項目をソースログと突合し、不一致を全件直して数値一致させる（対象: docs/readme_claims.csv、draft 7a2a9b） | [
{
"agent_id": "planner",
"name": "Toa",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Minato",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "計画を置きます。README主張5項目をソースログと突合し、不一致を全件直して数値一致させる。担当は coder。次に進めます。",
"tool": null,
"artifact_refs": [
"a1"
],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "作業結果の初版です。README値を5... | [
{
"id": "a1",
"kind": "log",
"path": "logs/run_log.json",
"content_hash": "sha256:63866653ff32b03e0d05add6e1d9adbf1f44e82391ccfa0847581a73578aa44e",
"summary": "ソースログ（toy）",
"content": "{\"dup_rate\": 0.934, \"frac_open\": 0.817, \"keep_rate\": 0.144, \"toy_mse\": 0.274, \"toy_mse_hard\": 0.... | {
"success": true,
"metric": {
"name": "readme_mismatches",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 2,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 9,
"subtype": "mse_digit_shift",
"bank_id": "bank_train_v1",
"template_family_id": "doc_align/mse_digit_shift",
"attempt": 0,
"designed_negative": false,
"final_a... |
collab-train-000035 | 20,420,035 | train | ja | synth_ablation | 目標: `rho` の1因子表（7点・toy）を作り、c1<c2<…かつ cap=0.632 以下を数値assertで通す（対象: ablations/rho.csv、draft 364588） | [
{
"agent_id": "planner",
"name": "Mei",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Emi",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Hina",
"brief": "誤り・リークを指摘。編集とツール不可"
},
{
"agent_id": "verifier",
"name": "... | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "段取りはこうします。`rho` の1因子表（7点・toy）を作り、c1<c2<…かつ cap=0.632 以下を数値assertで通す。担当は coder、critic、verifier。根拠はartifactにあります。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "ask",
"utt... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/rho.csv",
"content_hash": "sha256:36458887a08e8cef81777536e8bc6560bbd7865d3248e05c29d8e372d10fd1b2",
"summary": "初版",
"content": "factor,value,toy_accept\nrho,0.181,0.39699999999999996\nrho,0.3,0.305\nrho,0.419,0.407\nrho,0.538,0.249\nrho,0.6... | {
"success": true,
"metric": {
"name": "strict_monotone_caps",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 4,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 9,
"subtype": "rho_monotone",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/rho_monotone",
"attempt": 0,
"designed_negative": false,
"final_ar... |
collab-train-000036 | 20,420,036 | train | ja | protocol_review | 狙いはプロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする（対象: protocols/eval_protocol.md、draft f68f35） | [
{
"agent_id": "planner",
"name": "Hina",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Yui",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Emi",
"brief": "誤り・リークを指摘。編集とツール不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "全体像から入ります。プロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする。担当は coder、critic。次に進めます。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "ask",
"utterance": "質問があります。成功は誰が判定？記録に残します。... | [
{
"id": "a1",
"kind": "protocol",
"path": "protocols/eval_protocol.md",
"content_hash": "sha256:f68f3543450a39744d9a086bc666d9b13c69ef5cade449a3162af0e3f710b576",
"summary": "初版",
"content": "CALIB: 20960001-20960023 (thresholds only)\nTEST: 20960023-20960127\nHOLD: 20960133-20960160 (sealed... | {
"success": true,
"metric": {
"name": "protocol_violations",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 14,
"subtype": "seed_overlap",
"bank_id": "bank_train_v1",
"template_family_id": "protocol_review/seed_overlap",
"attempt": 0,
"designed_negative": false,
"final_... |
collab-train-000037 | 20,420,037 | train | mix | doc_align | 目標: README主張4項目をソースログと突合し、不一致を全件直して数値一致させる（対象: docs/readme_claims.csv、draft a4ac43） | [
{
"agent_id": "planner",
"name": "Aoi",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Yui",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "verifier",
"name": "Minato",
"brief": "検証器の実行のみ。編集不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "役割を割り振ります。README主張4項目をソースログと突合し、不一致を全件直して数値一致させる。担当は coder、verifier。根拠はartifactにあります。",
"tool": null,
"artifact_refs": [
"a1"
],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance":... | [
{
"id": "a1",
"kind": "log",
"path": "logs/run_log.json",
"content_hash": "sha256:05612c834b9e4a29f170127cbed706fa93315f02ee8a750a1914b36c28bd0ea9",
"summary": "ソースログ（toy）",
"content": "{\"archive_bytes\": 690007, \"keep_rate\": 0.356, \"toy_mse_hard\": 0.292, \"toy_mse_soft\": 0.855}"
},
... | {
"success": true,
"metric": {
"name": "readme_mismatches",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 10,
"subtype": "kib_calc",
"bank_id": "bank_train_v1",
"template_family_id": "doc_align/kib_calc",
"attempt": 0,
"designed_negative": false,
"final_artifact": "a3... |
collab-train-000038 | 20,420,038 | train | ja | synth_ablation | 成功条件は`accept_floor` の1因子表（6点・toy）を作り、c1<c2<…かつ cap=0.613 以下を数値assertで通す（対象: ablations/accept_floor.csv、draft ba7921） | [
{
"agent_id": "planner",
"name": "Rin",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Toa",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "verifier",
"name": "Tsumugi",
"brief": "検証器の実行のみ。編集不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "計画を置きます。`accept_floor` の1因子表（6点・toy）を作り、c1<c2<…かつ cap=0.613 以下を数値assertで通す。担当は coder、verifier。次に進めます。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance":... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/accept_floor.csv",
"content_hash": "sha256:ba7921a771474fe3e2eea4353b40d883dbaa8b30d330fccf0a66221a4d571b63",
"summary": "初版",
"content": "factor,value,toy_accept\naccept_floor,0.199,0.369\naccept_floor,0.326,0.165\naccept_floor,0.454,-0.044\... | {
"success": false,
"metric": {
"name": "strict_monotone_caps",
"value": 0
},
"stopped_reason": "budget"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 7,
"subtype": "accept_positive",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/accept_positive",
"attempt": 0,
"designed_negative": true,
"fin... |
collab-train-000039 | 20,420,039 | train | ja | mqar_gen | 目標: MQAR縮小サンプル3件（keys=3, vocab=256）を生成し、集合演算検証の違反数を0にする（対象: mqar/samples.jsonl、draft 595540） | [
{
"agent_id": "planner",
"name": "Yui",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Rin",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Tsumugi",
"brief": "誤り・リークを指摘。編集とツール不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "段取りはこうします。MQAR縮小サンプル3件（keys=3, vocab=256）を生成し、集合演算検証の違反数を0にする。担当は coder、critic。問題なければ進めます。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "まず一案を置きます... | [
{
"id": "a1",
"kind": "code",
"path": "mqar/samples.jsonl",
"content_hash": "sha256:5955401dbc27fe56d60ecae74cec2009deb1f8f021218021db8a9b89f51c3db0",
"summary": "初版",
"content": "{\"kv\": [[0, 178], [1, 205], [2, 173]], \"query\": [2, 0], \"answer\": [173, 178]}\n{\"kv\": [[0, 206], [4, 167... | {
"success": true,
"metric": {
"name": "mqar_violations",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 6,
"subtype": "key_duplicate",
"bank_id": "bank_train_v1",
"template_family_id": "mqar_gen/key_duplicate",
"attempt": 0,
"designed_negative": false,
"final_artifa... |
collab-train-000040 | 20,420,040 | train | ja | protocol_review | 狙いはプロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする（対象: protocols/eval_protocol.md、draft 7de80a） | [
{
"agent_id": "planner",
"name": "Yui",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Rin",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "役割を割り振ります。プロトコル草案のリーク（帯交差・禁止パターン）を全て除去し、検証器の違反数を0にする。担当は coder。よろしくお願いします。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "reply",
"utterance": "叩き台です。帯を CALIB 20730001-2... | [
{
"id": "a1",
"kind": "protocol",
"path": "protocols/eval_protocol.md",
"content_hash": "sha256:7de80acb7c7bbb7833879a5fb23b659de4197cc798639c59155957f9e070f6c9",
"summary": "初版",
"content": "CALIB: 20730001-20730040 (thresholds only)\nTEST: 20730036-20730083\nHOLD: 20730087-20730141 (sealed... | {
"success": true,
"metric": {
"name": "protocol_violations",
"value": 0
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 2,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 6,
"subtype": "seed_overlap",
"bank_id": "bank_train_v1",
"template_family_id": "protocol_review/seed_overlap",
"attempt": 0,
"designed_negative": false,
"final_a... |
collab-train-000041 | 20,420,041 | train | ja | synth_ablation | 目標: `rho` の1因子表（5点・toy）を作り、c1<c2<…かつ cap=0.857 以下を数値assertで通す（対象: ablations/rho.csv、draft edabfe） | [
{
"agent_id": "planner",
"name": "Aoi",
"brief": "計画・分担・最終読み上げ。数値を自己申告しない"
},
{
"agent_id": "coder",
"name": "Toa",
"brief": "成果物を編集・テスト実行。秘密ファイル禁止"
},
{
"agent_id": "critic",
"name": "Minato",
"brief": "誤り・リークを指摘。編集とツール不可"
}
] | [
{
"t": 0,
"role": "planner",
"act": "plan",
"utterance": "段取りはこうします。`rho` の1因子表（5点・toy）を作り、c1<c2<…かつ cap=0.857 以下を数値assertで通す。担当は coder、critic。根拠はartifactにあります。",
"tool": null,
"artifact_refs": [],
"claims": []
},
{
"t": 1,
"role": "coder",
"act": "ask",
"utterance": ... | [
{
"id": "a1",
"kind": "table",
"path": "ablations/rho.csv",
"content_hash": "sha256:edabfe2193b7030fe2d91e2126531417d812ed85c5dd5651c31f398b36d68174",
"summary": "初版",
"content": "factor,value,toy_accept\nrho,0.061,0.46199999999999997\nrho,0.211,0.256\nrho,0.361,0.466\nrho,0.511,0.483\nrho,0... | {
"success": true,
"metric": {
"name": "strict_monotone_caps",
"value": 1
},
"stopped_reason": "done"
} | {
"max_turns": 16,
"n_agents": 3,
"generator": "collab-synth-v1",
"no_lm_sota_claim": true,
"quality_tag": "hq-v1",
"n_turns_target": 9,
"subtype": "rho_monotone",
"bank_id": "bank_train_v1",
"template_family_id": "synth_ablation/rho_monotone",
"attempt": 0,
"designed_negative": false,
"final_ar... |
End of preview. Expand in Data Studio
pseudo-collab-hq（疑似協調軌跡データセット・quality_tag=hq-v1）
planner / coder / critic / verifier の2〜4エージェントが、1つのタスクを「計画 → 作業 → 検査 → 批評 → 改訂 → 検証 → 確定」の順に進める合成軌跡です。実際のツールログではありません。
- スキーマ:
collab-traj-schema-v1（SCHEMA.md）。hq-v1 で meta を拡張しています。 - 品質仕様:
QUALITY_SPEC.md。実測値はQUALITY.mdにあります。 - 生成器:
collab-synth-v1。テンプレート＋手続き生成＋実行型の検証器で作っており、LLMは使っていません。
| split | 件数 | success率 | turns mean/var | seed帯 |
|---|---|---|---|---|
| train | 500 | 0.912 | 10.17/8.39 | 20420001–20420500 |
| test | 100 | 0.91 | 10.21/7.73 | 20430001–20430100 |
| holdout | 50 | 0.92 | 10.1/8.57 | 20440001–20440050 |
- task_type（各20%）: bugfix / synth_ablation / protocol_review / doc_align / mqar_gen。
- サブタイプ: タスクごとに6種（train 3 / test 2 / holdout 1）。split間で互いに素です（
meta.subtype、meta.template_family_id）。 - 言い回しバンク:
bank_train_v1/bank_test_v1/bank_hold_v1は非共有です。分割は CALIB seed 20410001 で固定しています。 - 分布: turns（6–7 / 8–10 / 11–13 / 14–16 = 20 / 40 / 25 / 15%）と agents（2 / 3 / 4 = 25 / 45 / 30%）は、seed帯内の固定割当表で決まります。目標分布との KL はほぼ0です。
- 成功判定:
outcome.successは検証器だけが決めます。検証器の内容は次のとおりです。- bugfix: コードを exec して6ケースを照合。
- synth_ablation: 厳密単調と cap の数値 assert。
- protocol_review: seed帯の交差と禁止パターン。
- doc_align: ソースログとの数値一致。
- mqar_gen: 集合演算による検証。
- 負例: 約9%を意図的な負例（改訂が不十分なケース）として入れています。
success=false、stopped_reason=budgetです。 - ゲート: G1–G10 は全件 PASS です。近傍重複は char-3gram Jaccard で 0.85 未満、grounded率は 1.0 です。
- 旧版: 旧ビルド（7ターン固定・4エージェント固定）は以前の HF リビジョンに残っています。
注意: 数値はすべて toy です。LMや協調の性能は主張しません。artifacts[].content は検証用のスキーマ拡張です。
再生成: python build_pseudo_collab.py（phrase_banks.py が必要）
- Downloads last month
- 49