yomi-chen commited on
Commit
81b1efd
·
verified ·
1 Parent(s): 082cb89

Clarify stricter asset judge prompt anchors and unchanged scoring interface

Browse files
EditableDesignBench.zip CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:1765f17ef23279bea574cfbc38f4da38a79aea97b5f2fe8b6aa9f75abf48fe8e
3
- size 523878134
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:93d1d7d58afc6cb323f76f54e63e37afe0991012dc67fed0f8c584993e1fe560
3
+ size 523878671
README.md CHANGED
@@ -83,10 +83,14 @@ The MLLM judge makes four independent 0–5 scoring requests per task:
83
  | Text | Instruction following (IF), text accuracy (TA), visual consistency (VC), layout preservation (LP) |
84
  | Assets | Instruction following (IF), asset-state accuracy (OA), visual coherence (VC), non-target layout/content preservation (LP) |
85
 
86
- Asset judges receive exactly two native-resolution full images (source and prediction), the original instruction and the operation. The current asset MLLM protocol is editabledesignbench-asset-mllm-v3; no detail crops or masks are read or supplied to the judge. They do not receive the reference target, hidden replacement material, or the editing model's name. Natural movement/scale variations are accepted where the instruction leaves them unspecified. Unrelated text and graphics remain protected, including inside rectangular editing regions.
87
 
88
  Asset results include per-dimension and per-operation scores and a joint qualification rate: all four dimensions must score at least 4. Its denominator is all selected tasks; missing model outputs count as unqualified. API failures remain pending and are retried rather than converted into model scores. Scores and coverage are reported together. Asset criteria and the joint rate are project-specific extensions, not an official TextEditBench asset protocol. Report the two tracks separately.
89
 
 
 
 
 
90
  ## Scoring protocol update (2026-11-04)
91
 
92
  Asset judging uses two full images. Protocol v3 additionally requires source/candidate/change evidence per criterion, explicitly distinguishes insertion from replacing another image, and flags conflicting factual claims across independent requests for review. Invalid or self-contradictory responses are retried; scores are never rescaled. Re-evaluate asset MLLM scores with the updated code; old and new protocols use separate cache identities. Task data, pixel metrics and text judging are unchanged.
 
83
  | Text | Instruction following (IF), text accuracy (TA), visual consistency (VC), layout preservation (LP) |
84
  | Assets | Instruction following (IF), asset-state accuracy (OA), visual coherence (VC), non-target layout/content preservation (LP) |
85
 
86
+ Asset judges receive exactly two native-resolution full images (source and prediction), the original instruction and the operation. The current asset MLLM protocol is editabledesignbench-asset-mllm-v4; no detail crops or masks are read or supplied to the judge. They do not receive the reference target, hidden replacement material, or the editing model's name. Natural movement/scale variations are accepted where the instruction leaves them unspecified. Unrelated text and graphics remain protected, including inside rectangular editing regions.
87
 
88
  Asset results include per-dimension and per-operation scores and a joint qualification rate: all four dimensions must score at least 4. Its denominator is all selected tasks; missing model outputs count as unqualified. API failures remain pending and are retried rather than converted into model scores. Scores and coverage are reported together. Asset criteria and the joint rate are project-specific extensions, not an official TextEditBench asset protocol. Report the two tracks separately.
89
 
90
+ ## Stricter prompt anchors (2026-11-05)
91
+
92
+ Asset protocol v4 refines the judge prompt: 5 requires a verified excellent result with no visible criterion-specific defect, 4 requires every essential property to hold with at most a barely noticeable cosmetic issue, and clear material failures belong at 3 or below. Explicit anchors address unintended resizing during movement, changed attributes, intrinsic clipping and occluded neighboring content. Correct edits can still receive 5; scores are not rescaled to a desired distribution. Inputs, response schema, parser, retry behavior and four independent requests remain unchanged. Calibration is diagnostic and never a required batch gate.
93
+
94
  ## Scoring protocol update (2026-11-04)
95
 
96
  Asset judging uses two full images. Protocol v3 additionally requires source/candidate/change evidence per criterion, explicitly distinguishes insertion from replacing another image, and flags conflicting factual claims across independent requests for review. Invalid or self-contradictory responses are retried; scores are never rescaled. Re-evaluate asset MLLM scores with the updated code; old and new protocols use separate cache identities. Task data, pixel metrics and text judging are unchanged.
assets/evaluation_protocol.json CHANGED
@@ -26,7 +26,7 @@
26
  "scripts/mllm_evaluate.py",
27
  "scripts/asset_judge_profile.py"
28
  ],
29
- "judge_version": "editabledesignbench-asset-mllm-v3",
30
  "criteria": {
31
  "IF": "Requested asset operation is completed on the intended unit.",
32
  "OA": "Requested asset state, category, count, attributes and visible identity are correct.",
@@ -62,8 +62,8 @@
62
  "official_TextEditBench_asset_metric": false,
63
  "model_results_included": false,
64
  "release_note": "The asset scorer is implemented in the associated evaluation repository; this dataset release contains no model predictions or model evaluation results.",
65
- "version": "editabledesignbench-asset-mllm-v3",
66
- "message_template_revision": "asset-evidence-full-image-v5",
67
  "views": [
68
  "full source",
69
  "full prediction"
@@ -103,6 +103,13 @@
103
  "evidence_field_definitions": {
104
  "target_match": "Whether the intended source unit/empty slot is identified and any actual edit concerns it; not requested final-state completion. An unchanged correctly identified target/empty slot is yes with operation not_done. A visible edit of another source unit/slot is no.",
105
  "operation": "Whether the requested operation on the intended unit is done, partial, not_done, wrong, or uncertain."
 
 
 
 
 
 
 
106
  }
107
  }
108
  }
 
26
  "scripts/mllm_evaluate.py",
27
  "scripts/asset_judge_profile.py"
28
  ],
29
+ "judge_version": "editabledesignbench-asset-mllm-v4",
30
  "criteria": {
31
  "IF": "Requested asset operation is completed on the intended unit.",
32
  "OA": "Requested asset state, category, count, attributes and visible identity are correct.",
 
62
  "official_TextEditBench_asset_metric": false,
63
  "model_results_included": false,
64
  "release_note": "The asset scorer is implemented in the associated evaluation repository; this dataset release contains no model predictions or model evaluation results.",
65
+ "version": "editabledesignbench-asset-mllm-v4",
66
+ "message_template_revision": "asset-strict-anchor-full-image-v6",
67
  "views": [
68
  "full source",
69
  "full prediction"
 
103
  "evidence_field_definitions": {
104
  "target_match": "Whether the intended source unit/empty slot is identified and any actual edit concerns it; not requested final-state completion. An unchanged correctly identified target/empty slot is yes with operation not_done. A visible edit of another source unit/slot is no.",
105
  "operation": "Whether the requested operation on the intended unit is done, partial, not_done, wrong, or uncertain."
106
+ },
107
+ "prompt_revision_note": "Only criterion-specific prompt anchors are stricter; image inputs, output schema, score parser, retry mechanism and independent four-request format are unchanged.",
108
+ "strict_anchor_examples": {
109
+ "IF": "Visible extra rescaling during movement or incomplete execution: at most 3.",
110
+ "OA": "Clearly changed semantic attribute, identity, proportions or intrinsic completeness: at most 3.",
111
+ "VC": "Substantial clipping of an intrinsic part: at most 2.",
112
+ "LP": "Meaningful new occlusion of neighboring readable text or graphics: at most 2."
113
  }
114
  }
115
  }