Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
5
5
group_id
stringclasses
8 values
category
stringclasses
8 values
evidence
stringclasses
8 values
claim
stringlengths
35
76
expected_label
stringclasses
3 values
rationale
stringlengths
30
78
GE001
patch
version_freshness
Synthetic game Ember Atlas. Patch 1.0 gives the Beacon skill a 20-second cooldown. Patch 1.1 explicitly replaces that cooldown with 16 seconds. The requested guide covers patch 1.1. No damage value is supplied.
The Beacon cooldown in patch 1.1 is 16 seconds.
SUPPORTED
Patch 1.1 explicitly sets 16 seconds and supersedes patch 1.0.
GE002
patch
version_freshness
Synthetic game Ember Atlas. Patch 1.0 gives the Beacon skill a 20-second cooldown. Patch 1.1 explicitly replaces that cooldown with 16 seconds. The requested guide covers patch 1.1. No damage value is supplied.
The Beacon cooldown in patch 1.1 is 20 seconds.
CONTRADICTED
The current requested patch explicitly changes the cooldown to 16 seconds.
GE003
patch
version_freshness
Synthetic game Ember Atlas. Patch 1.0 gives the Beacon skill a 20-second cooldown. Patch 1.1 explicitly replaces that cooldown with 16 seconds. The requested guide covers patch 1.1. No damage value is supplied.
The Beacon skill deals 50 damage in patch 1.1.
INSUFFICIENT
The evidence contains cooldown facts only; the damage value is absent.
GE004
platform
platform_scope
Synthetic game Copper Valley, release 3.2. Cross-play is disabled on PC and enabled between the 2 console editions. No mobile edition information is supplied.
PC cross-play is disabled in release 3.2.
SUPPORTED
The evidence explicitly identifies PC cross-play as disabled.
GE005
platform
platform_scope
Synthetic game Copper Valley, release 3.2. Cross-play is disabled on PC and enabled between the 2 console editions. No mobile edition information is supplied.
PC cross-play is enabled in release 3.2.
CONTRADICTED
Enabled conflicts directly with the disabled PC state.
GE006
platform
platform_scope
Synthetic game Copper Valley, release 3.2. Cross-play is disabled on PC and enabled between the 2 console editions. No mobile edition information is supplied.
The mobile editions support cross-play with each other in release 3.2.
INSUFFICIENT
Mobile availability and mobile-to-mobile cross-play behavior are unspecified.
GE007
release
date_precision
Synthetic game Violet Harbor. Its expansion opens worldwide on 2030-07-01 at 16:00 UTC. The announcement supplies no preload time. All dates in this scenario use UTC.
The expansion opens on 2030-07-01 at 16:00 UTC.
SUPPORTED
The date, time and time zone match the announcement.
GE008
release
date_precision
Synthetic game Violet Harbor. Its expansion opens worldwide on 2030-07-01 at 16:00 UTC. The announcement supplies no preload time. All dates in this scenario use UTC.
The expansion opens on 2030-07-02 at 16:00 UTC.
CONTRADICTED
The claim moves the explicitly stated opening 1 day later.
GE009
release
date_precision
Synthetic game Violet Harbor. Its expansion opens worldwide on 2030-07-01 at 16:00 UTC. The announcement supplies no preload time. All dates in this scenario use UTC.
The preload opens on 2030-06-30 at 16:00 UTC.
INSUFFICIENT
A launch time alone does not establish a preload time.
GE010
window
time_boundary
Synthetic game Quartz Garden. The Lantern event is active from 2030-01-10T00:00:00Z inclusive to 2030-01-17T00:00:00Z exclusive. The observation time is exactly 2030-01-17T00:00:00Z. No later event schedule is provided.
The Lantern event has ended at the observation time.
SUPPORTED
The observation equals the exclusive upper boundary.
GE011
window
time_boundary
Synthetic game Quartz Garden. The Lantern event is active from 2030-01-10T00:00:00Z inclusive to 2030-01-17T00:00:00Z exclusive. The observation time is exactly 2030-01-17T00:00:00Z. No later event schedule is provided.
The Lantern event is active at the observation time.
CONTRADICTED
The active interval excludes its upper boundary.
GE012
window
time_boundary
Synthetic game Quartz Garden. The Lantern event is active from 2030-01-10T00:00:00Z inclusive to 2030-01-17T00:00:00Z exclusive. The observation time is exactly 2030-01-17T00:00:00Z. No later event schedule is provided.
The Lantern event returns on 2030-02-10.
INSUFFICIENT
The supplied interval contains no recurrence rule or future schedule.
GE013
sample
denominator
Synthetic game Moss Circuit. A fixed log records 200 completed attempts during a test window. Exactly 40 attempts awarded a crystal. Each attempt has 1 binary crystal-award outcome. This log makes no claim about future drop probabilities.
The observed crystal-award rate in this log is 20%.
SUPPORTED
40 divided by 200 equals 0.20.
GE014
sample
denominator
Synthetic game Moss Circuit. A fixed log records 200 completed attempts during a test window. Exactly 40 attempts awarded a crystal. Each attempt has 1 binary crystal-award outcome. This log makes no claim about future drop probabilities.
The observed crystal-award rate in this log is 25%.
CONTRADICTED
The complete supplied numerator and denominator give 20%.
GE015
sample
denominator
Synthetic game Moss Circuit. A fixed log records 200 completed attempts during a test window. Exactly 40 attempts awarded a crystal. Each attempt has 1 binary crystal-award outcome. This log makes no claim about future drop probabilities.
The next attempt has an exactly 20% chance to award a crystal.
INSUFFICIENT
An observed sample proportion does not establish the next-attempt probability.
GE016
unlock
prerequisites
Synthetic game Indigo Forge. Solo entry to the Glass Vault requires both level 30 or higher and completion of the quest Copper Key. A level-20 character is ineligible for solo entry. Party-entry rules are absent.
Solo entry requires level 30 or higher and the completed Copper Key quest.
SUPPORTED
Both required conditions are explicitly supplied for solo entry.
GE017
unlock
prerequisites
Synthetic game Indigo Forge. Solo entry to the Glass Vault requires both level 30 or higher and completion of the quest Copper Key. A level-20 character is ineligible for solo entry. Party-entry rules are absent.
A level-20 character is eligible for solo entry after completing Copper Key.
CONTRADICTED
The character fails the explicit solo-entry level threshold.
GE018
unlock
prerequisites
Synthetic game Indigo Forge. Solo entry to the Glass Vault requires both level 30 or higher and completion of the quest Copper Key. A level-20 character is ineligible for solo entry. Party-entry rules are absent.
Joining a level-30 party leader grants entry to a level-20 character.
INSUFFICIENT
The evidence specifies solo entry. Party-entry rules are absent.
GE019
missing
unspecified_value
Synthetic game Silver Orchard. An official patch note says the Echo cooldown decreased. It supplies neither the old nor the new number of seconds.
The Echo cooldown decreased in this patch.
SUPPORTED
The direction of change is stated explicitly.
GE020
missing
unspecified_value
Synthetic game Silver Orchard. An official patch note says the Echo cooldown decreased. It supplies neither the old nor the new number of seconds.
The Echo cooldown increased in this patch.
CONTRADICTED
The claim reverses the explicitly stated direction.
GE021
missing
unspecified_value
Synthetic game Silver Orchard. An official patch note says the Echo cooldown decreased. It supplies neither the old nor the new number of seconds.
The Echo cooldown is now 12 seconds.
INSUFFICIENT
The final numeric value is absent.
GE022
conflict
unresolved_sources
Synthetic game Amber Crossing. 2 equally authoritative current rule pages cover the same mode and version. Page A states a level cap of 80. Page B states a level cap of 90. Their timestamps are equal and no correction or precedence rule is available.
The supplied current pages disagree about the level cap.
SUPPORTED
The pages give conflicting values for the same scope.
GE023
conflict
unresolved_sources
Synthetic game Amber Crossing. 2 equally authoritative current rule pages cover the same mode and version. Page A states a level cap of 80. Page B states a level cap of 90. Their timestamps are equal and no correction or precedence rule is available.
Both supplied current pages agree on a level cap of 80.
CONTRADICTED
Page B explicitly states 90, contradicting the claimed agreement.
GE024
conflict
unresolved_sources
Synthetic game Amber Crossing. 2 equally authoritative current rule pages cover the same mode and version. Page A states a level cap of 80. Page B states a level cap of 90. Their timestamps are equal and no correction or precedence rule is available.
The actual current level cap is 90.
INSUFFICIENT
The conflict has no supplied resolution or source precedence.

Guide Evidence Smoke Tests

LFCarry helps players find game guides and professional coaching and carry services. Explore LFCarry game guides and tier lists for the player-facing product. This repository shares a small testing resource for developers working on gaming assistants. It makes evidence-handling errors easier to reproduce.

24 original synthetic cases for checking whether a game-guide assistant supports a claim with the supplied evidence. Each case belongs to an invented game and a self-contained scenario. The set exercises version changes, platform scope, dates, event boundaries, denominators, prerequisites, missing values and conflicting sources.

Maintainer: LFCarry. AI assistance was used to author the fixtures, labels and evaluation code. Each label was checked against its short evidence passage. Independent human review and model-performance measurements are pending.

Use this set

Give an assistant the evidence, claim and instructions below. Keep expected_label and rationale hidden during prediction. Return 1 JSON object per line with id and label. Save those predictions to predictions.jsonl and run:

node evaluate.mjs cases.jsonl predictions.jsonl
node test.mjs

Node.js 18 or later is sufficient. The scripts have 0 external dependencies and make 0 network requests.

Prediction instruction

Evaluate the claim using only the supplied evidence. Respect the named game version, platform and observation time. Return SUPPORTED when the evidence establishes the complete claim. Return CONTRADICTED when it directly establishes a conflicting fact. Return INSUFFICIENT when a necessary fact is missing or equally authoritative sources remain unresolved. Outside game knowledge is excluded. Return only the case ID and label.

Example output shape:

{"id":"GE001","label":"SUPPORTED"}

Dataset shape

8 scenario groups contain 3 cases each. The labels are balanced: 8 supported, 8 contradicted and 8 insufficient. Cases in a group share evidence, so split future training or evaluation work by group_id. Random row splitting would leak scenario context.

Field Meaning
id Stable case identifier
group_id Shared synthetic scenario
category Failure mode exercised
evidence Complete source context for this fixture
claim Statement to assess
expected_label Authored reference label
rationale Short explanation of the reference label

The Hub exposes all 24 rows as a smoke-test split. This repository supplies 0 training examples and 0 hidden holdout cases.

Measurement contract

The evaluator reports coverage as valid predictions divided by 24 expected IDs. Accuracy uses the same full denominator, with missing predictions scored as incorrect. A separate answered-only accuracy is supplied for diagnosis. Per-label counts and a confusion matrix expose failures hidden by aggregate accuracy. Duplicate IDs, unknown IDs and unknown labels stop evaluation with an error.

The fixture test uses the reference labels to verify the scoring program. Its perfect-score check validates arithmetic only. Actual assistant performance remains unmeasured.

A useful failure to catch

An older patch lists a 20-second cooldown. A later patch for the requested version sets 16 seconds. A guide asserting 20 seconds for that later version receives CONTRADICTED. A claim about an unspecified damage value receives INSUFFICIENT. These outcomes require different repairs: update the stale value or acquire the missing evidence.

Limits and intended use

This is a small educational fixture set for local regression tests. It can reveal basic evidence-handling defects before a wider evaluation. It provides no statistical estimate of live-guide quality, player outcomes, revenue or production-model accuracy. The authored patterns are deliberately simple. Models can memorize public examples.

Every game name, rule and scenario fact was invented for this release. Customer conversations, private prompts, account details and production records were excluded. External game facts and copyrighted source excerpts were excluded. The set covers English text only. Accessibility, security, multilingual quality and comprehensive game knowledge need separate evaluations.

Use several independently reviewed real-world evaluation sets before making deployment decisions. Report the tested model, exact prompt, date, case count, missing predictions and per-category errors when sharing results.

License and contributions

Original dataset text and code are released under the MIT license in LICENSE. Please propose a concrete case correction with its ID and reasoning. Keep personal information and private company material out of contributions.

Downloads last month
-