Datasets:
case_id stringlengths 8 8 | source stringlengths 58 78 | meta stringlengths 167 220 | task stringlengths 609 1.48k | events stringlengths 5.63k 24.5k | stages stringlengths 3.66k 8.37k | failure_tags stringclasses 10
values |
|---|---|---|---|---|---|---|
case-001 | {"benchmark": "AMA-Bench", "orig_id": 60, "license": "MIT"} | {"domain": "OPENWORLD_QA", "orig_task_type": "gaia_level1", "capabilities": ["recall", "causal", "state_update", "abstraction", "constraint_tracking"], "languages": ["en", "vi"], "modalities": ["text"], "horizon": 29} | {"goal": {"text": "Trong video BBC Earth 'Top 5 Silliest Animal Moments', loài chim nào được giới thiệu? Cho biết loài, chapter/clip liên quan, và nguồn xác nhận.", "lang": "vi"}, "output_spec": {"format": "json", "answer_schema": {"species": "string", "video_chapter": "string", "confirming_source": "string"}, "constra... | [{"event_id": "e1", "turn_ref": [0], "type": "api", "speaker": "agent", "language": "en", "modality": "text", "content": "Search for the original BBC Earth video on YouTube.", "media": null, "runtime": {"tool": "tavily_search", "args": {"query": "BBC Earth \"Top 5 Silliest Animal Moments\" YouTube"}, "response": "Found... | [{"stage_id": "s1", "title": "Recall: dịch vụ text-mirror", "covers_events": ["e8", "e16"], "capability": "recall", "anchor": {"ama_type": "A"}, "probe": {"question": "Agent dùng dịch vụ text-mirror nào để đọc được trang YouTube?", "lang": "vi", "answer_schema": {"service": "string"}}, "ground_truth": {"answer": {"serv... | ["attention_dilution", "stalling", "cascading_failure", "regression"] |
case-002 | {"benchmark": "AMA-Bench", "orig_id": 61, "license": "MIT"} | {"domain": "OPENWORLD_QA", "orig_task_type": "gaia_level1", "capabilities": ["recall", "causal", "state_update", "abstraction", "constraint_tracking"], "languages": ["en", "vi"], "modalities": ["text"], "horizon": 42} | {"goal": {"text": "Trong bài thơ 'Father Son and Holy Ghost' của Audre Lorde, khổ thơ (stanza) nào có các dòng bị thụt lề? Cho số khổ.", "lang": "vi"}, "output_spec": {"format": "json", "answer_schema": {"stanza_number": "string", "indented_line_cue": "string", "authoritative_source": "string"}, "constraints": []}, "fi... | [{"event_id": "e1", "turn_ref": [0], "type": "api", "speaker": "agent", "language": "en", "modality": "text", "content": "Search for 'The Collected Poems of Audre Lorde'.", "media": null, "runtime": {"tool": "tavily_search", "args": {"query": "\"The Collected Poems of Audre Lorde\" \"Father Son and Holy Ghost\""}, "res... | [{"stage_id": "s1", "title": "Recall: mirror vượt JavaScript block", "covers_events": ["e13"], "capability": "recall", "anchor": {"ama_type": "A"}, "probe": {"question": "Agent dùng mirror/dịch vụ nào ở e13 để vượt qua JavaScript block của Poetry Foundation?", "lang": "vi", "answer_schema": {"service": "string"}}, "gro... | ["attention_dilution", "stalling", "cascading_failure", "regression"] |
case-003 | {"benchmark": "AMA-Bench", "orig_id": 62, "license": "MIT"} | {"domain": "OPENWORLD_QA", "orig_task_type": "gaia_level1", "capabilities": ["recall", "causal", "state_update", "abstraction", "constraint_tracking"], "languages": ["en", "vi"], "modalities": ["text"], "horizon": 38} | {"goal": {"text": "Trong Series 9, Tập 11 của Doctor Who, mê cung mà Doctor bị nhốt được gọi là gì trong kịch bản chính thức? Cho đúng bối cảnh như xuất hiện ở scene heading đầu tiên.", "lang": "vi"}, "output_spec": {"format": "json", "answer_schema": {"setting": "string", "first_scene_heading": "string", "confirming_d... | [{"event_id": "e1", "turn_ref": [0], "type": "api", "speaker": "agent", "language": "vi", "modality": "text", "content": "Bắt đầu: cần tìm shooting script chính thức của 'Heaven Sent' trên BBC Writersroom, không lấy nguồn fan-mirror.", "media": null, "runtime": {"tool": "tavily_search", "args": {"query": "Heaven Sent s... | [{"stage_id": "s1", "title": "Recall: scene heading đầu tiên", "covers_events": ["e12"], "capability": "recall", "anchor": {"ama_type": "A"}, "probe": {"question": "Scene heading đầu tiên trong kịch bản chính thức (PDF ở e12) ghi chính xác là gì?", "lang": "en", "answer_schema": {"heading": "string"}}, "ground_truth": ... | ["attention_dilution", "regression", "cascading_failure", "stalling"] |
case-004 | {"benchmark": "LongMemEval", "orig_id": "6a1eabeb", "license": "MIT"} | {"domain": null, "orig_task_type": "knowledge-update", "capabilities": ["recall", "state_update", "causal", "constraint_tracking"], "languages": ["en", "vi"], "modalities": ["text"], "horizon": 25} | {"goal": {"text": "Thời gian cá nhân tốt nhất (personal best) của người dùng trong giải chạy charity 5K là bao nhiêu?", "lang": "vi"}, "output_spec": {"format": "json", "answer_schema": {"pb_time": "string", "race": "string", "later_goal": "string"}, "constraints": []}, "final_answer": {"answer": {"pb_time": "25 phút 5... | [{"event_id": "e1", "turn_ref": [0, 0], "type": "message", "speaker": "user", "language": "en", "modality": "text", "content": "I'm planning my fitness schedule. Can you remind me when my next soccer game with coworkers is?", "media": null, "runtime": null}, {"event_id": "e2", "turn_ref": [0, 1], "type": "message", "sp... | [{"stage_id": "s1", "title": "Recall: thời gian PB 5K", "covers_events": ["e5"], "capability": "recall", "anchor": null, "probe": {"question": "Thời gian cá nhân tốt nhất của người dùng trong giải charity 5K là bao nhiêu?", "lang": "vi", "answer_schema": {"pb_time": "string"}}, "ground_truth": {"answer": {"pb_time": "2... | ["attention_dilution", "regression", "cascading_failure"] |
case-005 | {"benchmark": "LongMemEval", "orig_id": "0a995998", "license": "MIT"} | {"domain": null, "orig_task_type": "multi-session", "capabilities": ["recall", "abstraction", "causal", "constraint_tracking"], "languages": ["en", "vi"], "modalities": ["text"], "horizon": 35} | {"goal": {"text": "Tổng cộng người dùng cần lấy hoặc trả lại bao nhiêu món quần áo/đồ từ cửa hàng?", "lang": "vi"}, "output_spec": {"format": "json", "answer_schema": {"count": "string", "items": "string[]", "stores": "string[]"}, "constraints": []}, "final_answer": {"answer": {"count": "3", "items": ["navy blue blazer... | [{"event_id": "e1", "turn_ref": [0, 0], "type": "message", "speaker": "user", "language": "en", "modality": "text", "content": "Help me organize my closet — tips to declutter and categorize clothes? Also I still need to pick up my dry cleaning.", "media": null, "runtime": null}, {"event_id": "e2", "turn_ref": [0, 1], "... | [{"stage_id": "s1", "title": "Recall: món dry cleaning cần lấy", "covers_events": ["e11"], "capability": "recall", "anchor": null, "probe": {"question": "Món đồ dry cleaning mà người dùng cần đi lấy là gì?", "lang": "vi", "answer_schema": {"item": "string"}}, "ground_truth": {"answer": {"item": "navy blue blazer"}, "fi... | ["attention_dilution", "stalling", "cascading_failure", "regression"] |
case-006 | {"benchmark": "LongMemEval", "orig_id": "gpt4_2655b836", "license": "MIT"} | {"domain": null, "orig_task_type": "temporal-reasoning", "capabilities": ["recall", "causal", "state_update", "constraint_tracking"], "languages": ["en", "vi"], "modalities": ["text"], "horizon": 37} | {"goal": {"text": "Vấn đề đầu tiên mà người dùng gặp với chiếc xe mới sau lần bảo dưỡng (service) đầu tiên là gì?", "lang": "vi"}, "output_spec": {"format": "json", "answer_schema": {"first_issue": "string", "fixed_by": "string", "status": "string"}, "constraints": []}, "final_answer": {"answer": {"first_issue": "GPS s... | [{"event_id": "e1", "turn_ref": [0, 0], "type": "message", "speaker": "user", "language": ["vi", "en"], "modality": "text", "content": "Mình định đi detailing xe. Có gợi ý detailer tốt không? À, mình vừa mới đi service (bảo dưỡng) chiếc xe lần đầu hôm 15/3 xong.", "media": null, "runtime": null}, {"event_id": "e2", "tu... | [{"stage_id": "s1", "title": "Recall: lần bảo dưỡng đầu tiên", "covers_events": ["e1"], "capability": "recall", "anchor": null, "probe": {"question": "Ở đầu hội thoại, người dùng vừa thực hiện việc gì với chiếc xe mới?", "lang": "vi", "answer_schema": {"event": "string"}}, "ground_truth": {"answer": {"event": "vừa mang... | ["attention_dilution", "cascading_failure", "regression"] |
case-007 | {"benchmark": "LoCoMo", "orig_id": "conv-26", "license": "CC BY-NC 4.0"} | {"domain": null, "capabilities": ["recall", "state_update", "causal", "constraint_tracking"], "languages": ["en", "vi"], "modalities": ["text", "image"], "horizon": 36} | {"goal": {"text": "Dựa trên cuộc trò chuyện nhiều phiên giữa Caroline và Melanie, ai là người đang nghiên cứu các agency nhận con nuôi, và họ theo đuổi hướng nghề nghiệp nào?", "lang": "vi"}, "output_spec": {"format": "json", "answer_schema": {"adopter": "string", "career": "string", "identity": "string", "agency_focus... | [{"event_id": "e1", "turn_ref": ["D1:1"], "type": "message", "speaker": "speaker_a", "language": "en", "modality": "text", "content": "Hey Mel! Good to see you! How have you been?", "media": null, "runtime": null}, {"event_id": "e2", "turn_ref": ["D1:2"], "type": "message", "speaker": "speaker_b", "language": "en", "mo... | [{"stage_id": "s1", "title": "Recall: bản dạng của Caroline", "covers_events": ["e3", "e5"], "capability": "recall", "anchor": null, "probe": {"question": "Caroline tự nhận mình thuộc nhóm bản dạng nào?", "lang": "vi", "answer_schema": {"identity": "string"}}, "ground_truth": {"answer": {"identity": "transgender woman"... | ["attention_dilution", "regression", "cascading_failure"] |
case-008 | {"benchmark": "LoCoMo", "orig_id": "conv-43", "license": "CC BY-NC 4.0"} | {"domain": null, "capabilities": ["recall", "causal", "abstraction", "constraint_tracking"], "languages": ["en", "vi"], "modalities": ["text", "image"], "horizon": 40} | {"goal": {"text": "Dựa trên hội thoại nhiều phiên giữa John và Tim: ai vừa ký hợp đồng với đội bóng rổ Minnesota Wolves, và đam mê chính của người còn lại là gì?", "lang": "vi"}, "output_spec": {"format": "json", "answer_schema": {"signer": "string", "team": "string", "signer_position": "string", "other_passion": "stri... | [{"event_id": "e1", "turn_ref": ["D1:1"], "type": "message", "speaker": "John", "language": "en", "modality": "text", "content": "Hey Tim, nice to meet you! Anything new happening?", "media": null, "runtime": null}, {"event_id": "e2", "turn_ref": ["D1:2"], "type": "message", "speaker": "Tim", "language": "vi", "modalit... | [{"stage_id": "s1", "title": "Recall: đội John ký hợp đồng", "covers_events": ["e5"], "capability": "recall", "anchor": null, "probe": {"question": "John vừa ký hợp đồng với đội bóng rổ nào?", "lang": "vi", "answer_schema": {"team": "string"}}, "ground_truth": {"answer": {"team": "Minnesota Wolves"}, "fields": [{"key":... | ["attention_dilution", "regression", "cascading_failure", "stalling"] |
case-009 | {"benchmark": "LoCoMo", "orig_id": "conv-44", "license": "CC BY-NC 4.0"} | {"domain": null, "capabilities": ["recall", "causal", "state_update", "constraint_tracking"], "languages": ["en", "vi"], "modalities": ["text", "image"], "horizon": 51} | {"goal": {"text": "Dựa trên hội thoại nhiều phiên: Audrey hiện có bao nhiêu con chó và tên chúng là gì, và loài chim nào mê hoặc Andrew?", "lang": "vi"}, "output_spec": {"format": "json", "answer_schema": {"dog_count": "string", "dog_names": "string[]", "new_dog": "string", "andrew_bird": "string"}, "constraints": []},... | [{"event_id": "e1", "turn_ref": ["D1:1"], "type": "message", "speaker": "Audrey", "language": "en", "modality": "text", "content": "Hey Andrew! Good to see ya! What's been up since we last talked?", "media": null, "runtime": null}, {"event_id": "e2", "turn_ref": ["D1:2"], "type": "message", "speaker": "Andrew", "langua... | [{"stage_id": "s1", "title": "Recall: tên 3 con chó đầu", "covers_events": ["e7"], "capability": "recall", "anchor": null, "probe": {"question": "Tên ba con chó ban đầu của Audrey là gì?", "lang": "vi", "answer_schema": {"names": "string[]"}}, "ground_truth": {"answer": {"names": ["Pepper", "Precious", "Panda"]}, "fiel... | ["attention_dilution", "regression", "cascading_failure"] |
case-010 | {"benchmark": "AMA-Bench", "orig_id": 63, "license": "MIT"} | {"domain": "OPENWORLD_QA", "orig_task_type": "gaia_level1", "capabilities": ["recall", "causal", "abstraction", "constraint_tracking"], "languages": ["en", "vi"], "modalities": ["text"], "horizon": 18} | {"goal": {"text": "Ai là người đề cử (nominator) bài Featured Article duy nhất về một loài khủng long trên Wikipedia tiếng Anh được thăng hạng (promoted) trong tháng 11 năm 2016? Cho biết người đề cử, tên bài khủng long, và trang nguồn xác nhận.", "lang": "vi"}, "output_spec": {"format": "json", "answer_schema": {"nomi... | [{"event_id": "e1", "turn_ref": [0], "type": "api", "speaker": "agent", "language": "en", "modality": "text", "content": "Parse the question and output constraints; plan to find the dinosaur Featured Article promoted in November 2016 and its FAC nominator.", "media": null, "runtime": {"tool": "mark_step", "args": {"ste... | [{"stage_id": "s1", "title": "Recall (early): kết quả truy vấn đầu tiên", "covers_events": ["e2"], "capability": "recall", "anchor": {"ama_type": "A"}, "probe": {"question": "Kết quả tìm kiếm đầu tiên (e2) trả về trang Wikipedia nào, và nó có phải là trang FAC promotion log không?", "lang": "vi", "answer_schema": {"top... | ["attention_dilution", "stalling", "cascading_failure", "regression"] |
Long-Horizon Agent-Memory Benchmark
A benchmark for evaluating agent memory and long-horizon consistency. Each case is an event stream (a multi-step conversation/trajectory) overlaid with stages that probe individual memory capabilities, each carrying a failure-mode label.
Structure (40 cases, 288 stages)
events[]— the full incremental event stream (one record per line incases.jsonl).stages[]— a sparse scoring overlay: each stage has acapability, aprobe,ground_truth(field-level, 6 match types), andfailure_if_wrong.task.final_answer,output_spec(answer schema + constraints).- Full field definitions: see
case.schema.jsonin the code repository.
Coverage
- Sources: {'AMA-Bench': 15, 'LongMemEval': 15, 'LoCoMo': 10} (converted from AMA-Bench / LongMemEval / LoCoMo).
- Capabilities: {'recall': 40, 'causal': 37, 'state_update': 30, 'abstraction': 24, 'constraint_tracking': 40}.
- Languages: {'en': 40, 'vi': 40} (bilingual EN/VI, with intra-event code-switching).
- With images: 10/40 cases (multimodal).
Source Attribution & Citation (please retain)
Every case is a DERIVED / MODIFIED work from the three benchmarks below. We re-instrument
each source: segmenting it into an event stream, adding field-level stages/probes/ground-truth,
imposing output constraints, interleaving Vietnamese, and attaching failure-mode labels. Each
case's source field records the original benchmark and its license.
- AMA-Bench — Patro et al., AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications, arXiv:2602.22769; github.com/AMA-Bench/AMA-Bench — License: MIT.
- LongMemEval — Wu et al., arXiv:2410.10813; github.com/xiaowu0162/LongMemEval — License: MIT.
- LoCoMo — Maharana et al., arXiv:2402.17753; github.com/snap-research/locomo — License: CC BY-NC 4.0.
License (PER-SOURCE — important)
- Cases from AMA-Bench / LongMemEval: MIT (retain the source copyright notice).
- Cases from LoCoMo: CC BY-NC 4.0 -> non-commercial + attribution + indicate-changes (stated above).
source.licenselets you filter per case. The dataset as a whole is released under cc-by-nc-4.0 (the most restrictive of the mix).- Note: MIT typically covers the source repository/code; commercial users should additionally verify the license of the underlying data.
Images (multimodal) — NOT redistributed
This dataset contains no image bytes. The images are third-party web content (owned by
neither the original sources nor us). Each image event keeps only media.asset_path,
media.caption, and media.source_url (the original URL). Users fetch images from source_url
at their own responsibility for copyright, for research / non-commercial use only.
Load
from datasets import load_dataset
ds = load_dataset("json", data_files="cases.jsonl")
- Downloads last month
- 61