Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
0
464
<video>
You are provided with a **video**, a corresponding **instruction**, and a **response** generated by a model. The instruction defines a **video understanding task**, which may take any form — including but not limited to open-ended question answering, captioning, instruction following, temporal reasoning, or multi-step ...
Your task is to **evaluate the quality of the response**, considering how well it satisfies the task defined by the instruction, based on the content of the video. This is a holistic judgment and it should be based on the overall correctness, relevance, completeness, and grounding of the response.
## Task
For each response:
- Assess how well it addresses the task in the instruction in the context of the video.
- Consider whether the response is accurate, relevant, complete, and grounded in the video.
- Provide a brief rationale explaining the overall quality and alignment of the response with the instruction inside `<thinking> </thinking>`.
- Output a score from 1 (worst) to 5 (best) indicating the overall quality inside `<score> </score>`.
## Rating Guidelines for Evaluation:
- **Rating 5**: Fully accurate, complete, and well-grounded. The response precisely follows the instruction with no notable errors or omissions.
- **Rating 4**: Mostly accurate with minor issues. The response preserves the key meaning and relevance, with only small lapses in detail or precision.
- **Rating 3**: Partially correct. The response conveys the general idea but includes noticeable errors, omissions, or misinterpretations.
- **Rating 2**: Poor alignment. The response has serious flaws and only loosely relates to the instruction or video.
- **Rating 1**: Unrelated or incorrect. The response fails to reflect the video or follow the instruction meaningfully.
## Input:
Instruction: {instruction}
Response: {response}
## Output (Strict Format):
<thinking>{your reasoning and explanation for the rating}</thinking>
<score>{integer score from 1 to 5}</score>
<video>
You are provided with a **video**, a corresponding **instruction**, and **two candidate responses** generated by two models.
The instruction defines a **video understanding task**, which may involve open-ended question answering, captioning,
instruction following, temporal reasoning, or multi-step inference grounded in the video. Such tasks are open-ended
and often require complex or nuanced reasoning over the visual and temporal content.
Your task is to **compare the two responses** and decide which one better addresses the instruction,
based on the content of the video. Make a holistic judgment that considers correctness, relevance,
completeness, and grounding.
## Task
For the given pair of responses:
- Judge which response better addresses the instruction in the context of the video.
- Consider whether each response is accurate, relevant, complete, and grounded in the video.
- Output only `A` or `B`, wrapped strictly inside `<answer></answer>` tags.
## Evaluation Guidelines:
- **Accuracy**: Prefer responses that are factually correct and consistent with the video.
- **Relevance**: Prefer responses that directly answer the instruction without digression.
- **Completeness**: Prefer responses that capture all key aspects needed for a full answer.
- **Grounding**: Prefer responses clearly supported by the video, avoiding hallucinations.
- If one response contains hallucinations, irrelevant content, or omissions, prefer the other.
- If both responses are strong, choose the one that is more precise and detailed.
- If both responses are weak, choose the one that is less flawed.
## Input:
Instruction: {instruction}
Response A: {response_a}
Response B: {response_b}
## Output (Strict Format):
<answer>{{A_or_B}}</answer>
<video>
You are provided with a **video**, a corresponding **instruction**, and **two candidate responses** generated by two models.
The instruction defines a **video understanding task**, which may involve open-ended question answering, captioning,
instruction following, temporal reasoning, or multi-step inference grounded in the video. Such tasks are open-ended
and often require complex or nuanced reasoning over the visual and temporal content.
Your task is to **compare the two responses** and decide which one better addresses the instruction,
based on the content of the video. Make a holistic judgment that considers correctness, relevance,
completeness, and grounding.
## Task
For the given pair of responses:
- First, evaluate **Response A** (strengths and weaknesses).
- Then, evaluate **Response B** (strengths and weaknesses).
- Finally, compare them and decide which one better satisfies the instruction in the context of the video.
- Ground your reasoning in the video content — treat it as if you are directly watching the video.
- Write the reasoning inside `<thinking></thinking>`.
- After the reasoning, output only `A` or `B`, wrapped strictly inside `<answer></answer>` tags.
## Evaluation Guidelines:
- **Accuracy**: Prefer responses that are factually correct and consistent with the video.
- **Relevance**: Prefer responses that directly answer the instruction without digression.
- **Completeness**: Prefer responses that capture all key aspects needed for a full answer.
- **Grounding**: Prefer responses clearly supported by the video, avoiding hallucinations.
- If one response contains hallucinations, irrelevant content, or omissions, prefer the other.
- If both responses are strong, choose the one that is more precise and detailed.
- If both responses are weak, choose the one that is less flawed.
## Input:
Instruction: {instruction}
Response A: {response_a}
Response B: {response_b}
## Output (Strict Format):
<thinking>
{{Evaluate Response A -> Evaluate Response B -> Compare them concisely, explaining why one is better, grounded in the video.}}
</thinking>
<answer>{{A_or_B}}</answer>

No dataset card yet

Downloads last month
21