text stringclasses 744
values |
|---|
2 |
8 |
5 |
51 |
46 |
79 |
86 |
87 |
95 |
99 |
112 |
115 |
123 |
127 |
136 |
144 |
145 |
142 |
146 |
148 |
149 |
147 |
151 |
150 |
152 |
153 |
154 |
156 |
158 |
159 |
163 |
161 |
162 |
164 |
165 |
166 |
167 |
168 |
170 |
172 |
174 |
173 |
176 |
179 |
178 |
183 |
182 |
184 |
185 |
186 |
187 |
191 |
192 |
195 |
194 |
197 |
198 |
199 |
200 |
201 |
202 |
203 |
207 |
210 |
214 |
215 |
217 |
218 |
219 |
220 |
221 |
226 |
228 |
227 |
229 |
230 |
232 |
233 |
2 |
5 |
51 |
79 |
86 |
87 |
95 |
99 |
112 |
115 |
123 |
127 |
136 |
142 |
144 |
145 |
146 |
147 |
148 |
149 |
150 |
151 |
YAML Metadata Warning:The task_categories "text-scoring" is not in the official list: text-classification, token-classification, table-question-answering, question-answering, zero-shot-classification, translation, summarization, feature-extraction, text-generation, fill-mask, sentence-similarity, text-to-speech, text-to-audio, automatic-speech-recognition, audio-to-audio, audio-classification, audio-text-to-text, voice-activity-detection, depth-estimation, image-classification, object-detection, image-segmentation, text-to-image, image-to-text, image-to-image, image-to-video, unconditional-image-generation, video-classification, reinforcement-learning, robotics, tabular-classification, tabular-regression, tabular-to-text, table-to-text, multiple-choice, text-ranking, text-retrieval, time-series-forecasting, text-to-video, image-text-to-text, image-text-to-image, image-text-to-video, visual-question-answering, document-question-answering, zero-shot-image-classification, graph-ml, mask-generation, zero-shot-object-detection, text-to-3d, image-to-3d, image-feature-extraction, video-text-to-text, keypoint-detection, visual-document-retrieval, any-to-any, video-to-video, other
SciTaste:跨学科科研创新性评测基准
一套面向大语言模型(LLM)的跨学科科研创新性评测基准,覆盖物理、数学、化学、生物学、医学、金融学 6 个学科,基于统一的 V6 七维度创新性指标,支持 4 种评测模式。
项目目标
- 系统性量化大语言模型在科研创新性评估任务上的能力
- 提供一套覆盖多学科、具有明确 GT(Ground Truth)的黄金评测集
- 提供多种评测模式(纯 QA / 原生搜索 / Agent + 工具 / Agent + 原生搜索),用于对比不同工具接入方式下的模型能力差异
- 用机器学习方法(随机森林/梯度提升)反推 V6 七维度权重,与人工设计权重形成互证
核心成果
- 6×1000 候选题目池(物理 + 数学 + 化学 + 生物学 + 医学 + 金融学)
- 6×50 黄金评测集(精选 300 篇,V6 七维度 GT 完整)
- 4 种评测模式 × 7+ 主流模型:qwen-3.8-max、qwen3.7-plus、qwen3.7-max、claude-fable-5、claude-opus-4-8、gpt-5.6-sol、gemini-3.5-flash
- V6 七维度创新性指数(citation_impact 29.5%、journal_impact 15.3%、conceptual_innovation 12.9%、knowledge_breadth 11.8%、semantic_novelty 11.6%、team_collaboration 10.1%、disruptiveness 8.9%)
- 随机森林权重分析:citation_impact 的 R² = 0.7359,证明其余六维可较好预测引用影响力
目录结构
科研创新性/
├── 独立论文部分/ # 物理八子领域 GT + 多学科 6×1000 候选池构建
├── 最终评测/ # 黄金 300 题 + 多模型评测结果 + OSS 上传包
├── 20260706-FrontierScience最新数据/ # FrontierScience 原始多模态评测
├── 基线部分/ # 基线模型(引用分布、语义新颖性、聚类)
├── 机器学习参数分析/ # V6 权重的机器学习推导(随机森林/梯度提升)
├── 团队维度映射分析/ # team_collaboration 维度映射方案
├── 数据文档说明-评测一体化模型训练.md
└── 评测一体化模型训练-QoderWork规划请求.md
主要子目录说明
独立论文部分/
存放所有"题目生成"相关的上游工作,包括:
- 物理八子领域(交叉/天体/应用/经典/计算/量子/凝聚态/高能)的原始数据爬取、创新性计算、CSV 输出
- 数学/化学/生物学/医学/金融学 5 个新学科的
*_innovation/目录,每个学科包含:crawler.py:Semantic Scholar / OpenAlex 爬取filter_and_supplement.py:学科分布、年份、期刊分区补充至 1000 篇compute_innovation.py+metrics.py:V6 七维度创新性计算run_*_innovation.py:一键生成 1000 题候选池
黄金评测集/新学科1000黄金题目集/:6 个学科各 1000 题的 JSONagent_rag_eval/:Agent vs RAG 对比评测与 V6 权重实验
最终评测/
核心产出目录,论文实验使用的最终数据都在这里。
黄金评测集/golden_evaluation_dataset_v5_final50_six_discipline_v2.json:6×50 黄金评测集(selected)+ 5700 后备(backup)agent_framework/:评测框架(LLM 客户端、Agent 引擎、9 个 ADS 工具、4 种模式配置、指标计算)eval50_*_sixdisc_v2/:每个模型四模式的原始*_results.jsonevaluation_results_report.md:所有模型的最终结果报告run_*.sh:各模型四模式评测启动脚本generate_results_report.py:报告生成脚本create_golden_evaluation_dataset.py:黄金 300 题生成入口compute_v6_spec.py:V6 七维度统一重算脚本
20260706-FrontierScience最新数据/
存放 2026 年 7 月 FrontierScience 原始多模态评测数据(Excel + JSON + 评测脚本),包括:
FS-多模态-*.xlsx/FS-纯文本-*.xlsx:各学科原始题目evaluate_frontier.py/evaluate_notools.py:Phase 2 双维度打分脚本evaluation_report_process_result.md:过程分/结果分报告score_calculation_methodology.md:评分方法论文档
基线部分/
基线对比实验:
calculate_novelty_*.py:语义新颖性基线计算fit_citation_*.py:引用分布拟合(log-normal/power-law)download_specter2.py:SPECTER2 模型下载
机器学习参数分析/
用随机森林 / 梯度提升反推 V6 七维度权重:
ml_parameter_analysis.py:主分析脚本V6维度权重机器学习推导方法.md结果总结.md/维度说明.md
团队维度映射分析/
team_collaboration 维度的作者数/国家数映射方案:
team_collaboration_mapping.py:主脚本mapping_lookup_table.csv:映射表
V6 七维度创新性指标
每篇论文的 Ground Truth 包含 7 个维度(0–100 分制),总分 = 加权和。
| 维度 | 英文键名 | 权重 | 计算类型 |
|---|---|---|---|
| A. 引用动态 | citation_impact |
29.5% | 引用数 + 百分位 |
| B. 参考文献多样性 | knowledge_breadth |
15.3% | PCA (ref_count, field/journal diversity, age variance) |
| C. 语义创新 | semantic_novelty |
12.9% | PCA (SS nearest neighbor / semantic distance) |
| D. 颠覆性 | disruptiveness |
11.8% | PCA (CD-index, cd_n_i) |
| E. 概念主题 | conceptual_innovation |
11.6% | PCA (core concept/topic scores, topic span) |
| F. 团队合作 | team_collaboration |
10.1% | PCA (country count, author count) |
| G. 期刊影响力 | journal_impact |
8.9% | IF rank + JCR + CAS |
公式:
ground_truth_score_100 = Σ(dim_i × weight_i) for i in 1..7
完整子参数与计算代码见:
最终评测/OSS_上传包_多学科创新性评测/reproducibility_details/compute_v6_spec.py最终评测/OSS_上传包_多学科创新性评测/reproducibility_details/可复现细节说明.md
评测模式
| 模式 | 工具 | 说明 |
|---|---|---|
pure_qa |
无 | 纯问答,无任何工具/搜索 |
native_search |
web_search | 模型原生搜索(如 OpenAI web_search) |
dimagent_v3 |
9 个 ADS 工具 | Agent 框架 + 结构化数据查询工具 |
dimagent_v3_native |
9 ADS + web_search | Agent 框架 + 结构化工具 + 原生搜索 |
9 个 ADS 工具(均只接受 bibcode 参数):
get_citation_dataget_journal_infoget_reference_listget_concept_dataget_author_dataget_citation_networkget_topic_dataget_disruptiveness_dataget_keyword_rarity
评测框架入口:
cd 最终评测/黄金评测集
python agent_framework/runs/run.py dimagent_v3 \
--model qwen3.7-max --resume --workers 1
数据格式
黄金评测集(JSON)
{
"dataset_name": "六学科 300 题黄金评测集 v2",
"total_selected": 300,
"total_backup": 5700,
"selected": [
{
"id": "https://openalex.org/Wxxxx",
"openalex_id": "https://openalex.org/Wxxxx",
"bibcode": "2020JCoPh.42309802F",
"title": "...",
"abstract": "...",
"year": 2020,
"doi": "10.1016/j.jcp.2020.109802",
"journal": "JCoPh",
"subcategory_id": "计算物理",
"discipline": "物理",
"ground_truth_score_100": 50.5,
"ground_truth_dimensions": {
"citation_impact": 32.24,
"journal_impact": 74.84,
"conceptual_innovation": 43.33,
"knowledge_breadth": 71.27,
"semantic_novelty": 64.57,
"team_collaboration": 19.51,
"disruptiveness": 68.31
},
"impact_factor": 3.8,
"partition": "Q1"
}
],
"backup": [ ... ]
}
评测结果(JSON)
每个模型每个模式一个 *_results.json,条目字段包括:
bibcode:论文标识status:success/failedmodel_prediction_score:模型预测总分model_prediction_dimensions:模型预测七维度分数ground_truth_score_100:GT 总分ground_truth_dimensions:GT 七维度tool_call_log:工具调用轨迹(仅 dimagent_v3 模式)
快速复现
1. 运行评测
cd 最终评测/黄金评测集
# 配置环境
export PYTHONUNBUFFERED=1
export EVAL_TOOL_BACKEND=openalex
export OPENALEX_MAP_FILE="openalex_bibcode_map_six_discipline.json"
export AIMUX_API_KEY="你的AIMUX密钥"
# 运行四模式评测(以 gemini-3.5-flash 为例)
bash run_gemini35flash_sixdisc_v2.sh
2. 生成报告
cd 最终评测/黄金评测集
python generate_results_report.py
# 输出:evaluation_results_report.md
3. 重算 V6 分数
cd 独立论文部分
python compute_v6_spec.py
4. 重建黄金评测集
cd 最终评测/黄金评测集
python create_golden_evaluation_dataset.py
# 输出:golden_evaluation_dataset_v5_final50_six_discipline_v2.json
当前评测结果摘要(最新)
| 模型 | Pure QA | Native Search | DimAgent-v3 | DimAgent + Native |
|---|---|---|---|---|
| qwen-3.8-max | 28.10 [E] | 31.60 [E] | 51.80 [D] | 50.30 [D] |
| qwen3.7-plus | 23.20 [E] | 20.80 [E] | 47.60 [D] | 48.10 [D] |
| qwen3.7-max | 20.10 [E] | 21.10 [E] | 47.70 [D] | 47.20 [D] |
| claude-fable-5 | 27.60 [E] | 32.70 [E] | 47.00 [D] | 39.30 [E] |
| claude-opus-4-8 | 25.50 [E] | 27.00 [E] | 34.50 [E] | 40.70 [D] |
| gpt-5.6-sol | 19.20 [F] | 19.50 [F] | 10.60 [F] | 28.70 [E] |
| gemini-3.5-flash | 16.40 [F] | 16.50 [F] | 17.70 [F] | 17.60 [F] |
档位:A (<5%)、B (5–10%)、C (10–20%)、D (20–30%)、E (30–50%)、F (≥50%),基于逐篇 sMAPE 均值。
已知问题与限制
- gemini-3.5-flash 的 native_search:187 条因
content_filter永久失败,属于 evamux/模型层内容过滤,无法通过调整 prompt 绕过 - claude-fable-5:部分生物/医学/化学题目因内容过滤返回空内容,失败数随模式递增(9/20/11/9)
- gpt-5.6-sol 的 dimagent_v3:工具模式下
temperature/reasoning_effort与 tools 共存报 400,已在llm_client.py中针对该模型禁用这两个参数 - 物理学科:原始八子领域数据集较大,重建物理 1000 题候选池需要依赖 CSV
技术栈
- 语言:Python 3.8+
- 主要依赖:
requests:API 调用scipy:统计检验numpy:数值计算scikit-learn:机器学习权重推导pandas:数据处理openai/evamux:LLM API 网关
- 数据源:
- NASA ADS:物理/天体论文元数据
- OpenAlex:多学科论文 + 引用 + 概念
- Semantic Scholar:语义新颖性特征
引用
如果你在论文或报告中引用本项目数据,请参考:
@misc{frontierscience2026,
title = {FrontierScience: Cross-disciplinary Scientific Innovation Evaluation Benchmark for Large Language Models},
author = {Gao, Xinyi and others},
year = {2026},
note = {6-discipline, V6-weighted, 4-mode evaluation framework},
}
许可
本项目采用 MIT 许可证。
- Downloads last month
- 86