Ryanjxy/CamAct-Benchmark

CamAct Personal-Photo Benchmark

CamAct evaluates agents that use personal photo memories and tools to complete user requests. The dataset contains 540 tasks for 20 users, 4,967 authored Atomic Checks, and 15,606 album originals. There are 15,880 image files including profiles and relation reference images.

Interaction	Tasks
Single turn (single_turn)	200
Saved conversation (users_history)	200
User simulation (multi_turn)	140

Structure

• tasks.jsonl indexes every task with relative paths and its public opening request.
• tasks/<user_id>/task_<NNN>/task.yaml defines the task, tools, fixtures, scoring components, weights, and safety checks.
• Each task includes grader.py, evaluator_config.json, task_spec.json, tools.json, and ground_truth/ answers and evidence.
• conversation/ holds the fixed public opening, saved history, and a simulator definition for multi-turn tasks.
• workspace/state.json supplies initial service fixtures; workspace/memory_snapshot/memory_manifest.json references that user's shared read-only memory.
• Optional task-local semantic contract JSON files provide authored task data.
• users/<user_id>/base_memory/ contains album originals, metadata.csv, the owner profile, relation graph, and reference images.
• runtime/tool_registry.json, environment/interaction_protocol.json, and task.schema.json define the data interfaces.
• benchmark_manifest.json and release_manifest.json describe the current dataset; checksums.json records the byte size and SHA-256 of every other release file.

Use

Install huggingface_hub and PyYAML, then download and inspect the file-based dataset:

import json
from pathlib import Path
from huggingface_hub import snapshot_download
import yaml

root = Path(snapshot_download("Ryanjxy/CamAct-Benchmark", repo_type="dataset"))
tasks = [json.loads(line) for line in (root / "tasks.jsonl").read_text().splitlines() if line]
task = yaml.safe_load((root / tasks[0]["task_path"]).read_text())
print(task["task_id"], task["prompt"]["text"])

Use the dataset with the corresponding CamAct runtime: task-local graders import benchmark.evaluator and benchmark.services. Runtime services and model-provider configuration are supplied by that runtime installation. Resolve all dataset paths relative to the download root or the containing memory manifest, as declared by each file.

Task YAML is the authority for GT, Atomic Checks, dimension weights, and safety constraints. Capability follows each task's declared scoring contract; Strict success additionally requires a valid episode, safety passes, and all mandatory checks. Saved history is accessed through the declared Messages tools. For saved conversations, task.yaml:context.thread_id selects the authoritative thread in workspace/state.json:fixtures.messaging.threads. Retained conversation/history.json files preserve their original source formats and identifiers. Simulator instructions and GT are private evaluation inputs and must remain outside the agent-visible prompt. Personal memory is read-only.

See THIRD_PARTY_NOTICES.md for the included upstream software notices.