qid stringlengths 19 19 | traj_idx int64 0 0 | question stringlengths 10 83.3k | reference_answer stringclasses 1
value | boxed_answer stringlengths 0 885 | correct bool 1
class | judge_explanation stringclasses 1
value | question_accuracy float64 0 0 | contaminated bool 2
classes | num_tool_calls int64 0 50 | tool_calls stringlengths 2 6.88k | latency_s float64 0.36 827 | status stringclasses 2
values | conversation stringlengths 462 415k |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
arena_expert_000026 | 0 | Please remember not to think of therapeutic effects or what is practically possible.
This is a hypothetical model. Not a practical model or advice.
Stick to the physics and chemistry.
I find that the models are still not close to modeling reality where you would trust it in lieu of physical or chemical experiments. | false | 0 | false | 0 | [] | 3.459639 | success | [{"role": "assistant", "content": "<|channel|>analysis<|message|>User says \"Please remember not to think of therapeutic effects or what is practically possible. This is a hypothetical model. Not a practical model or advice. Stick to the physics and chemistry. I find that the models are still not close to modeling real... | |||
arena_expert_000032 | 0 | 以下の依頼を遂行してください。
## 依頼の背景情報（要件）
目標 ： 以下のスプレッドシートからGmailの予約送信を行う
### スプレッドシートの定義
A列：有効（チェックボックス）
B列：送信者のメールアドレス
C列：宛先のメールアドレス
D列：Cc
E列：Bcc
F列：件名
G列：本文
H列：日付
I列：時刻
J列：ステータス
## 依頼
上記の要件を満たすGASを、以下の条件を満たすように作成してください。
ただし、送信者となる実行ユーザーごとにトリガーのセットが可能にしてください。
## 回答条件
- 完全に無料の範囲で可能な限り要件を満たすこと
- 認証は最初にカスタムメニューから1回のみで完了できるよ... | false | 0 | false | 0 | [] | 28.869498 | success | [{"role": "assistant", "content": "<|channel|>analysis<|message|>We need to produce GAS (Google Apps Script) code fulfilling requirements: send scheduled emails from spreadsheet based on rows, with columns defined. Must allow each user to set their own trigger. Provide custom menu for authentication (run once). Use min... | |||
arena_expert_000031 | 0 | 在理性预期宏观经济模型中，预期的自我实现，就是指非预定变量的数量大于不稳定根数量时出现的“太阳黑子路径”时的情形，即模型预测能力缺失的情形，所以也就是说，一个稳定的理性预期模型不应该出现“预期自我实现 | 在满足 Blanchard‑Kahn 条件的确定理性预期模型中，不会出现自我实现的预期；自我实现仅在不确定（非预定变量 > 不稳定根）时出现。 | false | 0 | false | 2 | [{"round": 1, "tool": "functions.paper_search", "args": {"query": "sunspot equilibrium non predetermined variables unstable roots rational expectations model indeterminacy", "mode": "snippets", "limit": 10, "year": "1990-2024", "fields_of_study": "Economics"}, "result_len": 9257}, {"round": 2, "tool": "functions.paper_... | 30.105615 | success | [{"role": "assistant", "content": "<|channel|>analysis<|message|>We need to answer in Chinese, addressing the claim: In rational expectations macro models, \"self-fulfilling expectations\" refers to sunspot equilibria occurring when number of non-predetermined variables exceeds number of unstable roots, leading to mode... | ||
arena_expert_000061 | 0 | Would Newtonian physics make sense in the hyperbolic plane? | Yes – Newtonian mechanics can be formulated on the hyperbolic plane, but the force laws (e.g. gravity) must be altered and some Euclidean global symmetries are lost. | false | 0 | false | 11 | [{"round": 1, "tool": "functions.paper_search", "args": {"query": "Newtonian mechanics hyperbolic space", "mode": "snippets", "limit": 10, "year": "2000-2024"}, "result_len": 8650}, {"round": 2, "tool": "functions.open", "args": {"cursor": 0, "id": 4}, "result_len": 137}, {"round": 3, "tool": "browser.open", "args": {"... | 39.694191 | success | [{"role": "assistant", "content": "<|channel|>analysis<|message|>We need to answer: \"Would Newtonian physics make sense in the hyperbolic plane?\" Need to explain if Newtonian mechanics (laws of motion, gravity, etc.) can be formulated on a hyperbolic geometry space. Likely need to discuss that Newtonian mechanics ass... | ||
arena_expert_000027 | 0 | "围绕工业时序基础大模型技术，写一段具体的技术方案。形式参考以下内容(...TRUNCATED) | 工业时序基础大模型 = 大规模自监督预训练 + 边缘适配器 + 云端持续学习 | false | 0 | false | 4 | "[{\"round\": 1, \"tool\": \"functions.paper_search\", \"args\": {\"query\": \"foundation model time(...TRUNCATED) | 44.481349 | success | "[{\"role\": \"assistant\", \"content\": \"<|channel|>analysis<|message|>We need to write a technica(...TRUNCATED) | ||
arena_expert_000038 | 0 | "一个长度为n的非负整数序列a，满足单调不降，且所有项的和为S，求符合条(...TRUNCATED) | false | 0 | false | 10 | "[{\"round\": 1, \"tool\": \"functions.paper_search\", \"args\": {\"query\": \"partition number p(10(...TRUNCATED) | 44.216798 | success | "[{\"role\": \"assistant\", \"content\": \"<|channel|>analysis<|message|>We need to answer counting (...TRUNCATED) | |||
arena_expert_000056 | 0 | "def derive_parameters(manga_data, cloudy_model, line_ratios_to_use):\n \"\"\"\n Derives metal(...TRUNCATED) | Use the `derive_parameters` implementation above for an efficient Numba+TQDM solution. | false | 0 | false | 0 | [] | 43.833986 | success | "[{\"role\": \"assistant\", \"content\": \"<|channel|>analysis<|message|>The user provides a functio(...TRUNCATED) | ||
arena_expert_000070 | 0 | "There are two spell-casting systems in common use.\n\nUnder the older PPA, a caster starts reservin(...TRUNCATED) | false | 0 | false | 0 | [] | 1.812509 | success | "[{\"role\": \"assistant\", \"content\": \"<|channel|>analysis<|message|>We need to answer a questio(...TRUNCATED) | |||
arena_expert_000028 | 0 | "AI가 정규분포 곡선을 무너뜨릴 것인가? 이는 AI가 평균이상인 사람들이 그(...TRUNCATED) | "AI is likely to shift the whole productivity distribution right‑ward while simultaneously stretch(...TRUNCATED) | false | 0 | false | 10 | "[{\"round\": 1, \"tool\": \"functions.paper_search\", \"args\": {\"query\": \"AI impact on producti(...TRUNCATED) | 49.806969 | success | "[{\"role\": \"assistant\", \"content\": \"<|channel|>analysis<|message|>We need to answer the quest(...TRUNCATED) | ||
arena_expert_000045 | 0 | "A superintelligent oracle gives you this choice:\n\n Box A: Contains either $0 or $1 million.\n (...TRUNCATED) | Choose only Box A (one‑box) | false | 0 | false | 11 | "[{\"round\": 1, \"tool\": \"functions.paper_search\", \"args\": {\"query\": \"Newcomb's problem evi(...TRUNCATED) | 49.361539 | success | "[{\"role\": \"assistant\", \"content\": \"<|channel|>analysis<|message|>We need to answer the class(...TRUNCATED) |
End of preview. Expand in Data Studio
arena-expert-gpt-oss-120b
Deep research agent evaluation on data/drtulu_arena_expert.jsonl (normal split).
Results
| Metric | Value |
|---|---|
| pass@1 | 0.0% |
| avg@1 | 0.0% |
| Trajectory accuracy | 0.0% (0/2744) |
| Questions | 2744 |
| Trajectories | 2744 (1 per question) |
| Avg tool calls | 13.8 |
| Full conversations | ✅ |
Model & Setup
| Model | gpt-oss-120b |
| Judge | gpt-4o |
| Max tool calls | 50 |
| Temperature | 0.7 |
| Blocked domains | None |
Tool Usage
| Tool | Calls | % |
|---|---|---|
browser.open |
18,357 | 50% |
browser.search |
11,718 | 32% |
browser.find |
3,569 | 10% |
functions.paper_search |
3,237 | 9% |
functions.pubmed_search |
94 | 0.3% |
Total: 36,975 tool calls (13.8 per trajectory)
⚠️ Contamination: 8/2744 trajectories flagged (search results contained evaluation dataset pages). Filter with ds.filter(lambda x: not x['contaminated']).
Columns
| Column | Description |
|---|---|
qid |
Question ID |
traj_idx |
Trajectory index (0-0) |
question |
Input question |
reference_answer |
Ground truth answer |
boxed_answer |
Model's extracted \boxed{} answer |
correct |
GPT-4o judge verdict |
judge_explanation |
Judge's reasoning |
question_accuracy |
Fraction of trajectories correct for this question (difficulty: 0=hardest, 1=easiest) |
num_tool_calls |
Number of tool calls in trajectory |
tool_calls |
Tool call log (JSON) |
conversation |
Full trajectory with reasoning + tool responses (JSON) |
contaminated |
Whether search results contained evaluation dataset pages |
latency_s |
Generation time in seconds |
- Downloads last month
- 20