Datasets:
26TCA
/
TCA-Bench

Dataset card Data Studio Files
xet
 Community
1
text
stringlengths
0
377
2. **Source binding**: The candidate attributes the sound to the correct source category (`Foreground character`, `Off-screen character`, `Environment`).
A match is **incorrect** if the sound is absent from the candidate, or if the source category is wrong.
**Output Format:**
Return a JSON object only. No extra text.
```json
{
"results": [
{"index": 0, "correct": true, "reason": "<one short sentence>"},
{"index": 1, "correct": false, "reason": "<one short sentence explaining the binding error or missing event>"}
]
}
```
Candidate Caption:
{{Candidate}}
Ground Truth Audio JSON:
{{Ground_Truth}}
### Evaluation Prompt
**Role:** Temporal Logic Evaluator.
**Task:** For *every* relation pair in the **GT Relations JSON**, verify whether the **Candidate Caption** adequately preserves the described cross-modal temporal relationship.
---
**Guiding Principles:**
1. **Type-boundary tolerance.** Simultaneous ↔ lightly-sequential mismatches are acceptable. Only mark Incorrect when the candidate uses **strong** contradicting markers (e.g., "then"/"after" vs GT simultaneous, or "at the same time" vs GT causal chain).
2. **Causal direction matters.** Implicit ordering (A described before B) is fine, but if the candidate **clearly reverses** the GT's A→B direction, mark Incorrect.
3. **Temporal signal required.** Both events must appear and share *some* temporal/contextual link (word order, proximity, grammatical structure). Mentioning both in completely unrelated passages is not sufficient.
4. **Flexible semantic matching.** GT events are **salient, coarse-grained** — exact wording is not required. Different words or detail levels for the same action/sound/dialogue count as a match (e.g., "walks away" ≈ "leaves the room"). Vague/generic descriptions that could refer to many things do not.
5. **Ignore character identity.** This stage tests temporal understanding only. Swapped character attributions do not affect the verdict — only event ordering/causality/simultaneity matters.
6. **Narrative prose tolerance.** Events in nearby sentences or the same paragraph share implicit temporal context. Do not require co-occurrence in one sentence, but events in completely unrelated sections lack connection.
---
**Evaluation Criteria:**
For each GT pair, assign one of three verdicts — **Correct**, **Incorrect**, or **Skipped**:
* **Correct:**
* Both events (or close semantic equivalents) are present in the candidate.
* There is some implicit or explicit temporal/contextual link between them.
* The causal or chronological direction is not clearly reversed.
* Minor type mismatches (simultaneous ↔ lightly sequential) are tolerated per Principle 1.
* **Incorrect** — when both events **are** identifiable in the candidate, but:
* The causal or chronological direction is **clearly reversed** (not just ambiguous — actually reversed).
* The candidate **explicitly contradicts** the temporal relationship type (e.g., uses "then" / "after" for a GT `AV_simultaneous` pair, or asserts simultaneity for a GT causal chain).
* The two events appear in the candidate but with **no temporal or contextual connection** between them (mentioned in completely separate, unrelated passages).
* **Skipped** — when one or both GT events are **absent** from the candidate (not mentioned at all, even in paraphrase, or described so vaguely that the specific event cannot be identified). These relations cannot be evaluated for temporal understanding from the given text. **Do NOT mark as Incorrect when the problem...
---
**Inputs:**
* **GT Relations:** {{Ground_Truth_Relations}}
* **Candidate:** {{Candidate}}
**Output Format:**
Return ONLY the ratio of correct to evaluated (correct + incorrect) relations, excluding skipped. Format: `correct/evaluated` (e.g., `4/5`). If all relations are skipped, return `0/0`. No explanations.
### Evaluation Prompt
**Role:** Temporal Logic Evaluator.
**Task:** For *every* relation pair in the **GT Relations JSON**, verify whether the **Candidate Caption** adequately preserves the described cross-modal temporal relationship.
---
**Guiding Principles:**
1. **Type-boundary tolerance.** Simultaneous ↔ lightly-sequential mismatches are acceptable. Only mark Incorrect when the candidate uses **strong** contradicting markers (e.g., "then"/"after" vs GT simultaneous, or "at the same time" vs GT causal chain).
2. **Causal direction matters.** Implicit ordering (A described before B) is fine, but if the candidate **clearly reverses** the GT's A→B direction, mark Incorrect.
3. **Temporal signal required.** Both events must appear and share *some* temporal/contextual link (word order, proximity, grammatical structure). Mentioning both in completely unrelated passages is not sufficient.
4. **Flexible semantic matching.** GT events are **salient, coarse-grained** — exact wording is not required. Different words or detail levels for the same action/sound/dialogue count as a match (e.g., "walks away" ≈ "leaves the room"). Vague/generic descriptions that could refer to many things do not.
5. **Ignore character identity.** This stage tests temporal understanding only. Swapped character attributions do not affect the verdict — only event ordering/causality/simultaneity matters.
6. **Narrative prose tolerance.** Events in nearby sentences or the same paragraph share implicit temporal context. Do not require co-occurrence in one sentence, but events in completely unrelated sections lack connection.
---
**Evaluation Criteria:**
For each GT pair, assign one of three verdicts — **Correct**, **Incorrect**, or **Skipped**:
* **Correct:**
* Both events (or close semantic equivalents) are present in the candidate.
* There is some implicit or explicit temporal/contextual link between them.
* The causal or chronological direction is not clearly reversed.