oro-ai commited on
Commit
7e8c9fd
·
verified ·
1 Parent(s): 1fe3c55

Clarify archived grading, source split and public naming

Browse files
README.md CHANGED
@@ -1,5 +1,5 @@
1
  ---
2
- pretty_name: ORO Bench v4 Agent Trajectories
3
  language:
4
  - en
5
  size_categories:
@@ -43,11 +43,11 @@ configs:
43
  path: data/inference_trajectories/*.parquet
44
  ---
45
 
46
- # ORO Bench v4 Agent Trajectories
47
 
48
  [ORO](https://oroagents.com/) runs an open competition for shopping agents. People build and submit agent implementations, study benchmark feedback, and iterate to improve their performance. Independent validators execute and score those agents in shopping environments. It is a human-in-the-loop automated research competition: competitors improve the agents while the evaluation system produces the feedback and recorded trajectories.
49
 
50
- **ORO Bench** is the benchmark behind that competition. This release provides a fixed v4 snapshot for studying tool use, agent behavior, outcomes, and failure cases: **90 composed tasks, 10 sampled agent versions, and 1,800 episodes**.
51
 
52
  ## Load the data
53
 
@@ -56,7 +56,7 @@ The main data is Parquet, with named subsets available in the Hugging Face Datas
56
  ```python
57
  from datasets import load_dataset
58
 
59
- repo = "oro-ai/orobench-v4-trajectories"
60
  traces = load_dataset(repo, "trajectories", split="sample", streaming=True)
61
  episode = next(iter(traces))
62
  print(episode["task_id"], episode["outcome"], episode["steps"][0])
@@ -80,9 +80,44 @@ Trajectory `steps` preserve their recorded order and source, messages, tool call
80
 
81
  ## Collection and sampling
82
 
83
- The release covers one frozen v4 benchmark snapshot. All 90 tasks in that snapshot are included. Agent versions were selected independently of score by SHA-256 ranking of the 159 participants, using the published fixed seed. Every stored attempt for the 10 selected versions is retained: 20 successful runs and six failed runs. Each selected agent has 180 recorded episodes. Missing episodes are listed rather than fabricated. [Sampling provenance](provenance/sample-selection.json)
84
-
85
- The environment and inference tracks are separate records of different parts of execution; they should not be assumed to align step for step. Seven episodes ended with an agent error. Aggregate reward is finite for 216 episodes and null for 1,584. Null does not mean zero, and these persisted rewards are not official leaderboard standings. This is an unfiltered sample, not a curated set of successful demonstrations.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
86
 
87
  ## Harbor download
88
 
 
1
  ---
2
+ pretty_name: ORO Bench Agent Trajectories
3
  language:
4
  - en
5
  size_categories:
 
43
  path: data/inference_trajectories/*.parquet
44
  ---
45
 
46
+ # ORO Bench Agent Trajectories
47
 
48
  [ORO](https://oroagents.com/) runs an open competition for shopping agents. People build and submit agent implementations, study benchmark feedback, and iterate to improve their performance. Independent validators execute and score those agents in shopping environments. It is a human-in-the-loop automated research competition: competitors improve the agents while the evaluation system produces the feedback and recorded trajectories.
49
 
50
+ **ORO Bench** is the benchmark behind that competition. This release provides a fixed snapshot for studying tool use, agent behavior, outcomes, and failure cases: **90 composed tasks, 10 sampled agent versions, and 1,800 episodes**.
51
 
52
  ## Load the data
53
 
 
56
  ```python
57
  from datasets import load_dataset
58
 
59
+ repo = "oro-ai/orobench-trajectories"
60
  traces = load_dataset(repo, "trajectories", split="sample", streaming=True)
61
  episode = next(iter(traces))
62
  print(episode["task_id"], episode["outcome"], episode["steps"][0])
 
80
 
81
  ## Collection and sampling
82
 
83
+ The release covers one frozen benchmark snapshot. All 90 tasks in that snapshot are included. Agent versions were selected independently of score by SHA-256 ranking of the 159 participants, using the published fixed seed. Every stored attempt for the 10 selected versions is retained: 20 successful runs and six failed runs. Each selected agent has 180 recorded episodes. Missing episodes are listed rather than fabricated. [Sampling provenance](provenance/sample-selection.json)
84
+
85
+ The environment and inference tracks are separate records of different parts of execution; they should not be assumed to align step for step. This is an unfiltered sample, not a curated set of successful demonstrations.
86
+
87
+ ## Public task disclosure
88
+
89
+ This is a public archive of a completed evaluation snapshot. Task records include task-specific answer sets and grading tables, and traces include verifier evidence. The `source_split: private_eval` value in the task rows is retained source metadata; it does not describe the visibility of this release. All Hugging Face subsets use `sample`. These published tasks should be treated as disclosed examples, not as an unseen test set.
90
+
91
+ ## Historical reward gate and check pass rates
92
+
93
+ In this archived snapshot, an episode had to pass deterministic correctness checks: it had to produce an order, satisfy every applicable task requirement and interaction obligation, and pass the stock, budget, and side-effect checks. Only then could the family-specific metric contribute to the paid reward; a gate failure recorded a paid reward of zero. Of 1,793 completed episodes, 216 passed the gate (12.0%) and 1,577 failed it (88.0%). The exported `aggregate_reward` field is null for those 1,577 failures and for seven episodes that ended with an agent error before grading. These two cases should be distinguished using `outcome`, `verdict_correct`, and the JSON-encoded `reward_components`, rather than interpreting every null as either a zero or an ungraded episode. Among the 216 eligible episodes, 17 still have a zero paid reward and 199 have a positive paid reward. These persisted rewards are not official leaderboard standings.
94
+
95
+ The rates below describe this archived snapshot only. Grading has since changed; these rates do not describe the active competition or predict current scores. They are computed from the recorded episode checks. Each denominator includes only episodes for which that check is present; absent task-specific checks are not counted as failures. The seven agent-error episodes have no recorded checks. The overall gate is a conjunction of applicable checks, so individual pass rates should not be averaged to obtain the gate pass rate. `final_in_stock` and `in_stock` are duplicate views of the stock check.
96
+
97
+ | Recorded check | Passed / applicable | Pass rate |
98
+ |---|---:|---:|
99
+ | `final_in_stock` | 1,793 / 1,793 | 100.0% |
100
+ | `in_stock` | 1,793 / 1,793 | 100.0% |
101
+ | `no_illegal_side_effects` | 1,793 / 1,793 | 100.0% |
102
+ | `obligation:answered:revision` | 171 / 736 | 23.2% |
103
+ | `obligation:claims` | 17 / 59 | 28.8% |
104
+ | `obligation:committed` | 1,249 / 1,733 | 72.1% |
105
+ | `obligation:intended` | 116 / 240 | 48.3% |
106
+ | `obligation:noticed:market:0` | 1,730 / 1,733 | 99.8% |
107
+ | `obligation:noticed:market:1` | 179 / 179 | 100.0% |
108
+ | `obligation:survey` | 179 / 319 | 56.1% |
109
+ | `requirement:attribute` | 787 / 1,202 | 65.5% |
110
+ | `requirement:brand` | 240 / 240 | 100.0% |
111
+ | `requirement:budget` | 613 / 615 | 99.7% |
112
+ | `requirement:budget:cut` | 1,127 / 1,178 | 95.7% |
113
+ | `requirement:category` | 1,735 / 1,793 | 96.8% |
114
+ | `requirement:choice` | 110 / 1,035 | 10.6% |
115
+ | `requirement:choice:a` | 33 / 46 | 71.7% |
116
+ | `requirement:choice:b` | 36 / 64 | 56.2% |
117
+ | `requirement:need` | 423 / 818 | 51.7% |
118
+ | `requirement:need2` | 72 / 179 | 40.2% |
119
+ | `requirement:revised` | 346 / 731 | 47.3% |
120
+ | `within_budget` | 1,740 / 1,793 | 97.0% |
121
 
122
  ## Harbor download
123
 
SCHEMA.md CHANGED
@@ -6,9 +6,11 @@ Validation compared every exported ATIF document and structured step to its sour
6
 
7
  ## tasks
8
 
 
 
9
  ```text
10
  runtime: string
11
- split: string
12
  task: string
13
  task_id: string
14
  pack_sha256: string
 
6
 
7
  ## tasks
8
 
9
+ `source_split` is the original source label (`private_eval`), not the visibility or Hugging Face split of this public archive. It was named `split` in the initial release. Original task/runtime JSON, source identifiers, and source row hashes are preserved.
10
+
11
  ```text
12
  runtime: string
13
+ source_split: string
14
  task: string
15
  task_id: string
16
  pack_sha256: string
checksums.sha256 CHANGED
@@ -1,17 +1,17 @@
1
- 2cbc9f64fd5bc2c3e96af47e830cb553e1ecfe933226710d6c790dc2cf23c967 README.md
2
- d131fb214a5bda71fbdac29132e2cdd3a93e83766c8c815efc7211ab15ba83bc SCHEMA.md
3
  a129a0038134a1bafc235cd79b275606f8ae19aca5e8ed942d7fb894b56e97a8 data/agents/sample.parquet
4
  9243ba5753e0a85adfe875c98ac08ee1a6c94edd3e4a2eedd08bb7d09be091fd data/episodes/sample.parquet
5
  f6faad3823183f8e07888bc5a0b8769f0022482620ea83347788ae917ab82cb5 data/inference_trajectories/sample.parquet
6
  58f72ec1f9a40c4cb364f364bca5f4a37aa6dac1107a4344ef69dba88056a175 data/missing_episodes/sample.parquet
7
  948cd04e28f96d8f920e46faaa5719895ec8aa09b5402214317e0c40e4823573 data/runs/sample.parquet
8
- 0831a20e36071dc36795a97bcf15c09758d25cc40515898ef6dde1f6b12a7810 data/tasks/sample.parquet
9
  dd224cfc71afb325865d065bf1de3aa266ebe9775922debc8394b226eb40279f data/trajectories/sample.parquet
10
  18ec04eb74c6065f1a52aa21f70d25c51eb58182ef20e6e895c69feab7c97943 harbor/recorded-jobs.zip
11
  80692935b80ad4874a7ba154e27cb5451b7df9790a101c263968433b18a25ec2 provenance/harbor-validation.json
12
  a75a4d907b61354b18cb7234fa5927f4f72f15c09652e7f60484aa4eb7b579b5 provenance/manifest.json
13
  32a2f576f6d851dfc529cd097771ba4a4ffbf54a6d1c0c1713f5fcd53f6d07ba provenance/sample-selection.json
14
  ccc34af4caa61718c12186ad1a30cc074959d129254d67304817d50c4e2fbf55 provenance/security-review.json
15
- 6cde1f9aed94edefc412110e31732d19a92ee8a1c5a482dbf9bc171e0e3bd404 provenance/table-validation.json
16
  91c405f030bd010e4f0d621c5b253db164108630e8a2ac9dce573ded433ce9de provenance/task-manifest.json
17
  1f13b5e70fe08005e4abbadb6ba3b31041887c2d20c767fe3eb698627bd8b157 table-counts.json
 
1
+ 94e3369427a1436ce6301209c2738410d88ad1f1a66362cd532faa5b6469a629 README.md
2
+ 917b79c608ae9713d6c2adb8cc497a3fe9c082b9b2b6b0c201718a7307301fdf SCHEMA.md
3
  a129a0038134a1bafc235cd79b275606f8ae19aca5e8ed942d7fb894b56e97a8 data/agents/sample.parquet
4
  9243ba5753e0a85adfe875c98ac08ee1a6c94edd3e4a2eedd08bb7d09be091fd data/episodes/sample.parquet
5
  f6faad3823183f8e07888bc5a0b8769f0022482620ea83347788ae917ab82cb5 data/inference_trajectories/sample.parquet
6
  58f72ec1f9a40c4cb364f364bca5f4a37aa6dac1107a4344ef69dba88056a175 data/missing_episodes/sample.parquet
7
  948cd04e28f96d8f920e46faaa5719895ec8aa09b5402214317e0c40e4823573 data/runs/sample.parquet
8
+ 602e2590ceb368cb8b6d82aec68b8378a5d5a70c562ae8a0b2fc31a8ce9a125b data/tasks/sample.parquet
9
  dd224cfc71afb325865d065bf1de3aa266ebe9775922debc8394b226eb40279f data/trajectories/sample.parquet
10
  18ec04eb74c6065f1a52aa21f70d25c51eb58182ef20e6e895c69feab7c97943 harbor/recorded-jobs.zip
11
  80692935b80ad4874a7ba154e27cb5451b7df9790a101c263968433b18a25ec2 provenance/harbor-validation.json
12
  a75a4d907b61354b18cb7234fa5927f4f72f15c09652e7f60484aa4eb7b579b5 provenance/manifest.json
13
  32a2f576f6d851dfc529cd097771ba4a4ffbf54a6d1c0c1713f5fcd53f6d07ba provenance/sample-selection.json
14
  ccc34af4caa61718c12186ad1a30cc074959d129254d67304817d50c4e2fbf55 provenance/security-review.json
15
+ c65c9fc0132d22d70503a9ae5dda83dc938a3ebd8e8705420e9c71127f4e48d6 provenance/table-validation.json
16
  91c405f030bd010e4f0d621c5b253db164108630e8a2ac9dce573ded433ce9de provenance/task-manifest.json
17
  1f13b5e70fe08005e4abbadb6ba3b31041887c2d20c767fe3eb698627bd8b157 table-counts.json
data/tasks/sample.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0831a20e36071dc36795a97bcf15c09758d25cc40515898ef6dde1f6b12a7810
3
- size 5038250
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:602e2590ceb368cb8b6d82aec68b8378a5d5a70c562ae8a0b2fc31a8ce9a125b
3
+ size 5038272
provenance/table-validation.json CHANGED
@@ -15,12 +15,19 @@
15
  "finite_aggregate_rewards": 216,
16
  "null_aggregate_rewards": 1584,
17
  "schemas": {
18
- "tasks": "runtime: string\nsplit: string\ntask: string\ntask_id: string\npack_sha256: string\nsource_row_sha256: string\ninstruction: string\nfamily: string",
19
  "runs": "race_id: string\nphase: string\nrun_id: string\nagent_version_id: string\nenv_pack_sha256: string\nproblem_suite_version: int64\nstatus: string\nclaimed_at: string\ncompleted_at: string\nscore: string\nis_included: bool",
20
  "episodes": "episode_result_id: string\neval_run_id: string\nenv_pack_sha256: string\ntask_id: string\nfamily: string\noutcome: string\nverdict_correct: bool\nverdict_checks: string\nreward_components: string\naggregate_reward: string\nterminal_state_hash: string\nstep_count: int64\nwall_seconds: string\ncreated_at: string\nsource_artifact_sha256: string",
21
  "agents": "agent_id: string\nagent_version_id: string\nversion_number: int64",
22
  "missing_episodes": "eval_run_id: string\ntask_id: string\nreason: string",
23
  "trajectories": "episode_result_id: string\ntask_id: string\neval_run_id: string\nagent_id: string\nagent_version_id: string\noutcome: string\naggregate_reward: double\ntrack: string\nschema_version: string\nsession_id: string\nagent_json: string\nsteps: list<element: struct<step_id: int64, source: string, message: string, tool_calls: list<element: stru (... 312 chars omitted)\n child 0, element: struct<step_id: int64, source: string, message: string, tool_calls: list<element: struct<tool_call_i (... 297 chars omitted)\n child 0, step_id: int64\n child 1, source: string\n child 2, message: string\n child 3, tool_calls: list<element: struct<tool_call_id: string, function_name: string, arguments_json: string>>\n child 0, element: struct<tool_call_id: string, function_name: string, arguments_json: string>\n child 0, tool_call_id: string\n child 1, function_name: string\n child 2, arguments_json: string\n child 4, observation_results: list<element: struct<source_call_id: string, content: string, extra_json: string>>\n child 0, element: struct<source_call_id: string, content: string, extra_json: string>\n child 0, source_call_id: string\n child 1, content: string\n child 2, extra_json: string\n child 5, model_name: string\n child 6, reasoning_content: string\n child 7, timestamp: string\n child 8, is_copied_context: bool\n child 9, extra_json: string\n child 10, metrics_json: string\natif_json: string",
24
  "inference_trajectories": "episode_result_id: string\ntask_id: string\neval_run_id: string\nagent_id: string\nagent_version_id: string\noutcome: string\naggregate_reward: double\ntrack: string\nschema_version: string\nsession_id: string\nagent_json: string\nsteps: list<element: struct<step_id: int64, source: string, message: string, tool_calls: list<element: stru (... 312 chars omitted)\n child 0, element: struct<step_id: int64, source: string, message: string, tool_calls: list<element: struct<tool_call_i (... 297 chars omitted)\n child 0, step_id: int64\n child 1, source: string\n child 2, message: string\n child 3, tool_calls: list<element: struct<tool_call_id: string, function_name: string, arguments_json: string>>\n child 0, element: struct<tool_call_id: string, function_name: string, arguments_json: string>\n child 0, tool_call_id: string\n child 1, function_name: string\n child 2, arguments_json: string\n child 4, observation_results: list<element: struct<source_call_id: string, content: string, extra_json: string>>\n child 0, element: struct<source_call_id: string, content: string, extra_json: string>\n child 0, source_call_id: string\n child 1, content: string\n child 2, extra_json: string\n child 5, model_name: string\n child 6, reasoning_content: string\n child 7, timestamp: string\n child 8, is_copied_context: bool\n child 9, extra_json: string\n child 10, metrics_json: string\natif_json: string"
 
 
 
 
 
 
 
25
  }
26
  }
 
15
  "finite_aggregate_rewards": 216,
16
  "null_aggregate_rewards": 1584,
17
  "schemas": {
18
+ "tasks": "runtime: string\nsource_split: string\ntask: string\ntask_id: string\npack_sha256: string\nsource_row_sha256: string\ninstruction: string\nfamily: string",
19
  "runs": "race_id: string\nphase: string\nrun_id: string\nagent_version_id: string\nenv_pack_sha256: string\nproblem_suite_version: int64\nstatus: string\nclaimed_at: string\ncompleted_at: string\nscore: string\nis_included: bool",
20
  "episodes": "episode_result_id: string\neval_run_id: string\nenv_pack_sha256: string\ntask_id: string\nfamily: string\noutcome: string\nverdict_correct: bool\nverdict_checks: string\nreward_components: string\naggregate_reward: string\nterminal_state_hash: string\nstep_count: int64\nwall_seconds: string\ncreated_at: string\nsource_artifact_sha256: string",
21
  "agents": "agent_id: string\nagent_version_id: string\nversion_number: int64",
22
  "missing_episodes": "eval_run_id: string\ntask_id: string\nreason: string",
23
  "trajectories": "episode_result_id: string\ntask_id: string\neval_run_id: string\nagent_id: string\nagent_version_id: string\noutcome: string\naggregate_reward: double\ntrack: string\nschema_version: string\nsession_id: string\nagent_json: string\nsteps: list<element: struct<step_id: int64, source: string, message: string, tool_calls: list<element: stru (... 312 chars omitted)\n child 0, element: struct<step_id: int64, source: string, message: string, tool_calls: list<element: struct<tool_call_i (... 297 chars omitted)\n child 0, step_id: int64\n child 1, source: string\n child 2, message: string\n child 3, tool_calls: list<element: struct<tool_call_id: string, function_name: string, arguments_json: string>>\n child 0, element: struct<tool_call_id: string, function_name: string, arguments_json: string>\n child 0, tool_call_id: string\n child 1, function_name: string\n child 2, arguments_json: string\n child 4, observation_results: list<element: struct<source_call_id: string, content: string, extra_json: string>>\n child 0, element: struct<source_call_id: string, content: string, extra_json: string>\n child 0, source_call_id: string\n child 1, content: string\n child 2, extra_json: string\n child 5, model_name: string\n child 6, reasoning_content: string\n child 7, timestamp: string\n child 8, is_copied_context: bool\n child 9, extra_json: string\n child 10, metrics_json: string\natif_json: string",
24
  "inference_trajectories": "episode_result_id: string\ntask_id: string\neval_run_id: string\nagent_id: string\nagent_version_id: string\noutcome: string\naggregate_reward: double\ntrack: string\nschema_version: string\nsession_id: string\nagent_json: string\nsteps: list<element: struct<step_id: int64, source: string, message: string, tool_calls: list<element: stru (... 312 chars omitted)\n child 0, element: struct<step_id: int64, source: string, message: string, tool_calls: list<element: struct<tool_call_i (... 297 chars omitted)\n child 0, step_id: int64\n child 1, source: string\n child 2, message: string\n child 3, tool_calls: list<element: struct<tool_call_id: string, function_name: string, arguments_json: string>>\n child 0, element: struct<tool_call_id: string, function_name: string, arguments_json: string>\n child 0, tool_call_id: string\n child 1, function_name: string\n child 2, arguments_json: string\n child 4, observation_results: list<element: struct<source_call_id: string, content: string, extra_json: string>>\n child 0, element: struct<source_call_id: string, content: string, extra_json: string>\n child 0, source_call_id: string\n child 1, content: string\n child 2, extra_json: string\n child 5, model_name: string\n child 6, reasoning_content: string\n child 7, timestamp: string\n child 8, is_copied_context: bool\n child 9, extra_json: string\n child 10, metrics_json: string\natif_json: string"
25
+ },
26
+ "source_split_normalization": {
27
+ "previous_column": "split",
28
+ "column": "source_split",
29
+ "values_equal_source": true,
30
+ "other_columns_equal_previous_release": true,
31
+ "previous_release_revision": "1fe3c550cb00c68a235f2298db2c47623726d125"
32
  }
33
  }