Request access to CoPlayBench

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

CoPlayBench is provided for non-commercial research only. The game videos remain the property of their original copyright holders and are included solely so that research models can be evaluated and trained on them; do not redistribute the videos or use them for any commercial purpose. The annotations are licensed under CC BY-NC 4.0.

Log in or Sign Up to review the conditions and access this dataset content.

๐ŸŽฎ CoPlayBench

Benchmark data for CoPlay: Learning When to Speak and What to Say as Game Companions.

๐Ÿ’ป Code (evaluation + training) ยท ๐ŸŒ Project page ยท ๐Ÿ“„ Paper (coming soon)

A game companion watches the player's screen and must decide both what to say and whether to speak at all. CoPlayBench evaluates both decisions in two complementary modes:

  • Reactive: the player asks a question and the model answers from the last 60 s of screen (16 frames as four 2ร—2 grids).
  • Proactive: after one opening request the model watches the whole session, looks at the screen every 2โ€“6 s and decides at every step whether to speak, what to say and what to remember. A required window counts as hit only if the model speaks inside it.

Every event was marked independently by three annotator models and kept only if at least two agreed. It carries a speaking window, a player question, must-mention points, forbidden statements, a tone and a reference reply.

Split Videos Hours Events Reactive questions Proactive events
Bench 818 97.7 5,153 9,103 (5,153 in-play + 3,950 aftermath; 7,037 need advice) 5,133 (4,045 required + 1,088 optional)
Train 1,635 197 10,048 SFT 12,914 ยท RL 3,551 SFT 27,843 steps ยท RL 6,960 steps

Games (Bench videos): Honor of Kings 100 ยท Dou Dizhu 100 ยท Mahjong 100 ยท UNO 99 ยท Go 100 ยท Plants vs. Zombies 2 69 ยท Carrot Fantasy 126 ยท Hearthstone 91 ยท Stardew Valley 33. Bench and Train share no video (and no match for Honor of Kings).

๐Ÿ“ฆ Layout

bench/
  reactive.jsonl                9,103 reactive questions (one per line)
  proactive.jsonl               818 proactive sessions (one per video)
  reactive_timing_flags.jsonl   366 questions whose reference reply may mention a fact that appears after the question
                                time (flagged only; they are scored like every other question)
annotations/
  bench/<Game>.jsonl            English ground truth of the 818 Bench videos, one video per line
  train/<Game>.jsonl            English ground truth of the 1,635 Train videos
videos/
  bench/<Game>/<vid>.mp4        first-person gameplay videos (~33 GB)
  train/<Game>/<vid>.mp4        (~66 GB)
training/
  split.json                    Train video -> sft | dev | rl
  gt_timing.jsonl               timing audit of the Train GT: when each fact stated by a reference appears on screen
  sft_train.json, sft_dev.json  LlamaFactory sharegpt data (40,757 / 2,072 samples); image paths are relative
  dataset_info.json             LlamaFactory dataset registry
  rl/{train,dev}_{reactive,proactive}.jsonl   GRPO rows in training order (prompt, reward metadata, label)

The images referenced by training/ are not stored here: build_sft_data.py and build_rl_data.py in the code repository render them deterministically from the videos with the benchmark's own frame code.

๐Ÿš€ Usage

# accept the terms on this page first, then
hf auth login
hf download OpenRaiser/CoPlayBench --repo-type dataset --local-dir data/CoPlayBench \
    --include "bench/*" "videos/bench/*"          # evaluation only
from datasets import load_dataset
reactive = load_dataset("OpenRaiser/CoPlayBench", "reactive", split="test")
proactive = load_dataset("OpenRaiser/CoPlayBench", "proactive", split="test")

Running models, the judge and the metrics: see the code repository.

๐Ÿ—‚๏ธ Fields

bench/reactive.jsonl

Field Description
id <vid>_<node>_inplay or <vid>_<node>_aftermath
game, game_name, board game key, display name, whether it is one of the four board / card games
vid, video, duration_sec video id, path inside this dataset, length in seconds
question_type in_play (asked during the event) or aftermath (asked after it ended)
t_sec, player_message, gt_emotion when the player speaks, what they say, their annotated emotion
window, window_len_sec the event's speaking window and the time left in it at t_sec
scenario deadline class for OnTime: urgent (2 s), strategic (5 s), emotional (10 s)
event_type, tier, objective, situation, outcome, response_kind event description (outcome is shown to the judge for aftermath questions only)
must_mention, must_not points a good reply must cover / statements it must not make
vibe, emotion_requirements, max_words_en expected tone, support requirements, word limit
reference_response reference reply
agreement, gt_source, translated_by number of annotators that marked the event (2 or 3), the annotator whose annotation was kept, the translating model

bench/proactive.jsonl

Field Description
id, game, game_name, vid, video, duration_sec session = one whole video
step_sec seconds between two looks at the screen: 2 (board games), 4, or 6 (Stardew Valley)
proactive_instruction the player's only line, said at the start
quiet_spans_policy how much small talk is acceptable outside key moments
events[] node_id, type, tier (Required / Optional), window, response_kind, trigger, objective, situation, must_mention, must_not, vibe, emotion_requirements, max_words_en, reference_response, needs_advice, agreement, gt_source, translated_by
timing audit valid_from (when the reference first becomes true on screen), window_orig (window before the audit), timing_facts, timing; must_mention_orig / situation_orig where the audit changed them
dropped_events[] events whose reference is never true inside their window; lines spoken there are neither hits nor off-window lines

annotations/<split>/<Game>.jsonl

One video per line: vid, game, game_name, video, duration_sec, persona, proactive_instruction, quiet_spans_policy, covered_spans, rule, gt_scope (and pov_hero for Honor of Kings), plus nodes[], the retained events. Each node has the event description (type, tier, fight_span, objective, situation, outcome, pov_involvement, evidence_frames, agreement, agree_models, gt_source, translated_by) and up to three parts, each with its own window, success_criteria (must_mention, must_not, vibe, emotion_requirements, max_words_en) and reference_response:

  • in_play: the in-play question (t_sec, text, user_emotion, response_kind);
  • aftermath: the question after the event (question = t_sec, text, user_emotion);
  • proactive: the moment to speak up unprompted (response_kind, trigger).

bench/*.jsonl are derived from annotations/bench/ (the proactive events after the timing audit).

๐Ÿ“œ License and terms

  • Annotations (everything except videos/): CC BY-NC 4.0.
  • Videos: copyright remains with the original owners. They are provided only for non-commercial research through this gated dataset; do not redistribute them.
  • Code: Apache-2.0, in the code repository.

โœ’๏ธ Citation

@misc{coplay2026,
  title  = {CoPlay: Learning When to Speak and What to Say as Game Companions},
  author = {CoPlay Team},
  year   = {2026},
  url    = {https://github.com/OpenRaiser/CoPlayer}
}
Downloads last month
38