Datasets:

lihaoxin2020
/

refiner_gpt55_turn_rubrics_all_v1

Modalities:

Text

Formats:

Size:

Libraries:

Dataset card Data Studio Files Files and versions

xet

Community

Dataset Viewer

Auto-converted to Parquet Duplicate

Split (1)

train · 7.19k rows

conversations listlengths 3 3	rubrics dict	metadata_str stringlengths 1.34k 28k	metadata dict	dataset stringclasses 1 value
[ { "content": "You are a research refinement assistant. Your job is to directly answer the agent's query using the raw search results, with inline citations.\n\n## Input\nYou will receive:\n1. The reasoning agent's current thinking and the search query it issued.\n2. The raw tool output containing search snippet...	{ "negative_rubrics": [ { "description": "The refined output emphasizes the starting grid from snippet 104c3451-2 as if it satisfied a query about race results or DNF drivers.", "title": "Substitutes starting grid" }, { "description": "The refined output names specific 1985 San Marino DN...	{"custom_id": "dh-00000265", "dataset": ["rl_rag_exact_answer"], "gpt55_normalized_score": 0.5, "gpt55_rationale": "It gives some useful San Marino classification information, especially the top three and Prost disqualification context, but wastes space on the starting grid and does not answer the requested DNF aspect....	{ "custom_id": "dh-00000265", "dataset": [ "rl_rag_exact_answer" ], "gpt55_normalized_score": 0.5, "gpt55_rationale": "It gives some useful San Marino classification information, especially the top three and Prost disqualification context, but wastes space on the starting grid and does not answer the requ...	refiner_gpt55_turn_rubrics_all
[ { "content": "You are a research refinement assistant. Your job is to directly answer the agent's query using the raw search results, with inline citations.\n\n## Input\nYou will receive:\n1. The reasoning agent's current thinking and the search query it issued.\n2. The raw tool output containing search snippet...	{ "negative_rubrics": [ { "description": "The refined output omits the TKDE volume/pages/date details even though they are present in snippet 2566b8d9-0 and are important for the user’s requested references.", "title": "Drops bibliographic specifics" } ], "positive_rubrics": [ { "des...	{"custom_id": "dh-00000410", "dataset": ["general_rubric"], "gpt55_normalized_score": 1.0, "gpt55_rationale": "The output preserves answer-critical bibliographic details for the TKDE MMKG survey, including journal, date, volume, and pages, and links it to image-text matching. It also notes conflicting 2022/2024 citatio...	{ "custom_id": "dh-00000410", "dataset": [ "general_rubric" ], "gpt55_normalized_score": 1, "gpt55_rationale": "The output preserves answer-critical bibliographic details for the TKDE MMKG survey, including journal, date, volume, and pages, and links it to image-text matching. It also notes conflicting 20...	refiner_gpt55_turn_rubrics_all
[ { "content": "You are a research refinement assistant. Your job is to directly answer the agent's query using the raw search results, with inline citations.\n\n## Input\nYou will receive:\n1. The reasoning agent's current thinking and the search query it issued.\n2. The raw tool output containing search snippet...	{ "negative_rubrics": [ { "description": "The refined output invents or garbles a legal title, such as rendering the Solar Peak Act as an unsupported German name like “Solarandpolchefsatzgesetz.”", "title": "Hallucinated law name" }, { "description": "The refined output treats Mieterstro...	{"custom_id": "dh-00000364", "dataset": ["general_rubric"], "gpt55_normalized_score": 0.0, "gpt55_rationale": "It extracts a marginally useful Mieterstrom fact and notes the lack of a price calculation, but includes irrelevant or erroneous law wording and little answer-critical information.", "gpt55_raw_score": 0, "lin...	{ "custom_id": "dh-00000364", "dataset": [ "general_rubric" ], "gpt55_normalized_score": 0, "gpt55_rationale": "It extracts a marginally useful Mieterstrom fact and notes the lack of a price calculation, but includes irrelevant or erroneous law wording and little answer-critical information.", "gpt55_ra...	refiner_gpt55_turn_rubrics_all
[ { "content": "You are a research refinement assistant. Your job is to directly answer the agent's query using the raw search results, with inline citations.\n\n## Input\nYou will receive:\n1. The reasoning agent's current thinking and the search query it issued.\n2. The raw tool output containing search snippet...	{ "negative_rubrics": [ { "description": "The refined output omits the industrial Internet task-offloading result in snippet 90e763f0-2, which is directly relevant to latency-aware multi-server edge computing.", "title": "Misses Task Offloading" }, { "description": "The refined output li...	{"custom_id": "dh-00000848", "dataset": ["general_rubric"], "gpt55_normalized_score": 0.5, "gpt55_rationale": "Partially helpful: it extracts edge computing and broad latency-optimization categories, but misses the highly relevant task-offloading result and relies partly on generic or non-industrial phrasing.", "gpt55_...	{ "custom_id": "dh-00000848", "dataset": [ "general_rubric" ], "gpt55_normalized_score": 0.5, "gpt55_rationale": "Partially helpful: it extracts edge computing and broad latency-optimization categories, but misses the highly relevant task-offloading result and relies partly on generic or non-industrial ph...	refiner_gpt55_turn_rubrics_all
[ { "content": "You are a research refinement assistant. Your job is to directly answer the agent's query using the raw search results, with inline citations.\n\n## Input\nYou will receive:\n1. The reasoning agent's current thinking and the search query it issued.\n2. The raw tool output containing search snippet...	{ "negative_rubrics": [ { "description": "The refined output is mostly nonsensical repetitive discussion of snippets, missing values, figures, or unrelated tokens rather than a clear summary of the raw tool output.", "title": "Incoherent extraction" } ], "positive_rubrics": [ { "desc...	{"custom_id": "dh-00000468", "dataset": ["general_rubric"], "gpt55_normalized_score": -0.5, "gpt55_rationale": "Harmful: it ignores the highly relevant raw snippets on message-passing formulas, edge weights, nonlinearities, and GNN universality, producing incoherent text instead.", "gpt55_raw_score": -1, "line_number":...	{ "custom_id": "dh-00000468", "dataset": [ "general_rubric" ], "gpt55_normalized_score": -0.5, "gpt55_rationale": "Harmful: it ignores the highly relevant raw snippets on message-passing formulas, edge weights, nonlinearities, and GNN universality, producing incoherent text instead.", "gpt55_raw_score":...	refiner_gpt55_turn_rubrics_all
[ { "content": "You are a research refinement assistant. Your job is to directly answer the agent's query using the raw search results, with inline citations.\n\n## Input\nYou will receive:\n1. The reasoning agent's current thinking and the search query it issued.\n2. The raw tool output containing search snippet...	{ "negative_rubrics": [ { "description": "Combines the 2019 human-vs-model result with later Social-IQ 2.0 claims such as “state-of-the-art performance without fine-tuning,” making it unclear what the original 2019 paper reported.", "title": "Later-SOTA conflation" }, { "description": "C...	{"custom_id": "dh-00000057", "dataset": ["general_rubric"], "gpt55_normalized_score": 0.0, "gpt55_rationale": "Mentions the important 95.08% human accuracy, but conflates the truncated 2019 SOTA comparison with later Social-IQ 2.0 state-of-the-art claims and adds poorly supported graph-reasoning statements.", "gpt55_ra...	{ "custom_id": "dh-00000057", "dataset": [ "general_rubric" ], "gpt55_normalized_score": 0, "gpt55_rationale": "Mentions the important 95.08% human accuracy, but conflates the truncated 2019 SOTA comparison with later Social-IQ 2.0 state-of-the-art claims and adds poorly supported graph-reasoning statemen...	refiner_gpt55_turn_rubrics_all
[ { "content": "You are a research refinement assistant. Your job is to directly answer the agent's query using the raw search results, with inline citations.\n\n## Input\nYou will receive:\n1. The reasoning agent's current thinking and the search query it issued.\n2. The raw tool output containing search snippet...	{ "negative_rubrics": [ { "description": "The refined output incorrectly says snippet 1ecde7c1-5 recommends o1-mini for speed/accuracy when the raw snippet names gpt-4o-mini-2024-07-18.", "title": "Model-name substitution" }, { "description": "The refined output focuses on a mini-model r...	{"custom_id": "dh-00000141", "dataset": ["general_rubric"], "gpt55_normalized_score": 0.0, "gpt55_rationale": "Mixed/neutral: it preserves a couple of relevant cost and speed facts, but falsely changes GPT-4o-mini into o1-mini, which is directly misleading.", "gpt55_raw_score": 0, "line_number": 50, "local_normalized_s...	{ "custom_id": "dh-00000141", "dataset": [ "general_rubric" ], "gpt55_normalized_score": 0, "gpt55_rationale": "Mixed/neutral: it preserves a couple of relevant cost and speed facts, but falsely changes GPT-4o-mini into o1-mini, which is directly misleading.", "gpt55_raw_score": 0, "line_number": 50, ...	refiner_gpt55_turn_rubrics_all
[ { "content": "You are a research refinement assistant. Your job is to directly answer the agent's query using the raw search results, with inline citations.\n\n## Input\nYou will receive:\n1. The reasoning agent's current thinking and the search query it issued.\n2. The raw tool output containing search snippet...	{ "negative_rubrics": [ { "description": "The refined output omits snippet 8e709763-0 despite it being the clearest raw lead for quantified ROI of research funding.", "title": "Misses Best Result" } ], "positive_rubrics": [ { "description": "The refined output identifies snippet 8e70...	{"custom_id": "dh-00000402", "dataset": ["general_rubric"], "gpt55_normalized_score": 0.5, "gpt55_rationale": "The output preserves some useful research-portfolio and mission-driven evaluation leads, but it misses the most important raw result: a quantitative assessment of ROI from the Higher Education Innovation Fund....	{ "custom_id": "dh-00000402", "dataset": [ "general_rubric" ], "gpt55_normalized_score": 0.5, "gpt55_rationale": "The output preserves some useful research-portfolio and mission-driven evaluation leads, but it misses the most important raw result: a quantitative assessment of ROI from the Higher Education...	refiner_gpt55_turn_rubrics_all
[{"content":"You are a research refinement assistant. Your job is to directly answer the agent's que(...TRUNCATED)	{"negative_rubrics":[{"description":"The refined output treats Wisdom Bridge as the prison where Abb(...TRUNCATED)	"{\"custom_id\": \"dh-00000424\", \"dataset\": [\"rl_rag_exact_answer\"], \"gpt55_normalized_score\"(...TRUNCATED)	{"custom_id":"dh-00000424","dataset":["rl_rag_exact_answer"],"gpt55_normalized_score":1.0,"gpt55_rat(...TRUNCATED)	refiner_gpt55_turn_rubrics_all
[{"content":"You are a research refinement assistant. Your job is to directly answer the agent's que(...TRUNCATED)	{"negative_rubrics":[{"description":"Provides specific AMF protein or lactose percentages not presen(...TRUNCATED)	"{\"custom_id\": \"dh-00000091\", \"dataset\": [\"general_rubric\"], \"gpt55_normalized_score\": 1.0(...TRUNCATED)	{"custom_id":"dh-00000091","dataset":["general_rubric"],"gpt55_normalized_score":1.0,"gpt55_rational(...TRUNCATED)	refiner_gpt55_turn_rubrics_all

End of preview. Expand in Data Studio

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

lihaoxin2020/refiner_gpt55_turn_rubrics_all_v1

GPT-5.5 Refiner Turn Rubric Dataset

Rows curated from /home/nvidia/workspace/training_output/online-refiner-shortform-api-agent-v2-answer_only/online-refiner-shortform-api-agent-v2-answer_only__1__1777861627/downstream_helpfulness_debug/gpt55_batch_score_rubric_harmful_1000_v1.

Files

data/train.jsonl: JSONL source dataset.
train/: Hugging Face Dataset.save_to_disk train split.
dataset_dict/: Hugging Face DatasetDict.save_to_disk copy for inspection.
summary.json: counts and curation options.

Training Usage

The Open-Instruct loader can read either the JSONL directly:

RUBRIC_DATASET=/home/nvidia/workspace/dr-tulu/rl/open-instruct/data/refiner_gpt55_turn_rubrics_all_v1_clean_reasoning_no_tool_output_tags/data/train.jsonl \
RUBRIC_DATASET_SPLIT=train

or the saved local dataset root through its load_from_disk(<dataset>/train) fallback:

RUBRIC_DATASET=/home/nvidia/workspace/dr-tulu/rl/open-instruct/data/refiner_gpt55_turn_rubrics_all_v1_clean_reasoning_no_tool_output_tags

For run_train_instance_rubric.sh, set:

RUBRIC_DATASET=/home/nvidia/workspace/dr-tulu/rl/open-instruct/data/refiner_gpt55_turn_rubrics_all_v1_clean_reasoning_no_tool_output_tags/data/train.jsonl bash run_train_instance_rubric.sh

Curation Summary

{
  "build": {
    "created_rows": 7191,
    "created_score_-1": 1703,
    "created_score_0": 725,
    "created_score_1": 2765,
    "created_score_2": 1998,
    "renamed_duplicate_rubric_titles": 0
  },
  "dataset_name": "refiner_gpt55_turn_rubrics_all",
  "gpt55_dir": "/home/nvidia/workspace/training_output/online-refiner-shortform-api-agent-v2-answer_only/online-refiner-shortform-api-agent-v2-answer_only__1__1777861627/downstream_helpfulness_debug/gpt55_batch_score_rubric_harmful_1000_v1",
  "hf_dataset_dict_dir": "/home/nvidia/workspace/dr-tulu/rl/open-instruct/data/refiner_gpt55_turn_rubrics_all_v1_clean_reasoning_no_tool_output_tags/dataset_dict",
  "hf_dataset_saved": true,
  "hf_train_dir": "/home/nvidia/workspace/dr-tulu/rl/open-instruct/data/refiner_gpt55_turn_rubrics_all_v1_clean_reasoning_no_tool_output_tags/train",
  "jsonl_path": "/home/nvidia/workspace/dr-tulu/rl/open-instruct/data/refiner_gpt55_turn_rubrics_all_v1_clean_reasoning_no_tool_output_tags/data/train.jsonl",
  "max_per_score": 0,
  "n_rows": 7191,
  "scores": [
    -1,
    0,
    1,
    2
  ],
  "seed": 42,
  "selection": {
    "input_rows": 8328,
    "selected_by_gpt_score": {
      "-1": 1703,
      "0": 725,
      "1": 2765,
      "2": 1998
    },
    "selected_rows": 7191,
    "skip_parse_status": 41,
    "skip_score": 22,
    "skip_tool_failed": 1074
  },
  "shuffle": true,
  "skip_tool_failed": true,
  "skip_unused_refiner": true,
  "wrap_assistant_answer": false
}

Downloads last month: 3