fastino/fast-decisions

Dataset

Fast Decisions

5 commits

updated Sep 24, 2026

See the code

README

Fast Decisions

The classification suite behind GLiNER2.5-Decide. Seventeen domains, one file each. Every row is a document plus the decisions a product has to make: the task name, the candidate labels, and the gold label. One call can carry several heads. Single-label heads have one gold string. Multi-label heads list every label that applies.

This release is the development split: 100 examples per domain, 1,700 rows. The test split, 300 per domain, is held out. The numbers below are that held-out split. Do not report a score computed on the files in this repo as the benchmark.

This suite is not a general-purpose exam. There is no reasoning trace and no open answer. It is operational decisions: customer and banking intent, travel and clinic requests, review sentiment, document type, email and ticket routing, human handoff, and ordinal or yes/no gates.

Schema

Every row has this shape:

{
  "input": "<free text>",
  "output": {
    "classifications": [
      {
        "task": "<head name>",
        "true_label": [
          "<gold label(s)>"
        ],
        "labels": [
          "<label>",
          "..."
        ],
        "multi_label": false
      }
    ]
  }
}

true_label is always a list of strings, even for single-label heads.
labels is the full candidate set for that head — pass it to your model as-is; do not shuffle out the gold label.
Rows are classification only. There is no entity field.
Ordinal heads (0–5, 0–10) use string digits, not integers.

Domains

File	Domain	Group	Heads	Multi-label heads	Rows
`support_intent.jsonl`	Commerce support intent	commerce	1	—	100
`support_topic.jsonl`	Commerce support topic	commerce	1	—	100
`document_type.jsonl`	Document type	records	1	—	100
`review_sentiment.jsonl`	Review sentiment	reviews	1	—	100
`agent_handoff.jsonl`	Assistant handoff	assistants	1	—	100
`email_triage.jsonl`	Email triage	workplace	4	—	100
`ticket_route.jsonl`	Commerce ticket route	commerce	3	—	100
`product_feedback.jsonl`	Product feedback	commerce	2	`product_area`	100
`banking_intent.jsonl`	Banking intent	finance	1	—	100
`clinic_request.jsonl`	Clinic request	health	2	—	100
`travel_request.jsonl`	Travel request	travel	1	—	100
`news_topic.jsonl`	News topic	news	1	—	100
`paper_field.jsonl`	Paper field	science	1	—	100
`sports_recap.jsonl`	Sports recap	sports	3	—	100
`restaurant_review.jsonl`	Restaurant review	food	2	`aspects`	100
`benefits_request.jsonl`	Benefits request	civic	2	—	100
`screen_tags.jsonl`	Screen tags	entertainment	2	`genres`	100

Total: 1,700 rows across 17 files, 100 rows each.

Task heads and label sets

Full label inventory per head. Use these exact strings as the candidate set at inference time.

support_intent — Commerce support intent (commerce)

intent: order_status, refund_request, cancel_subscription, update_payment, login_problem, password_reset, shipping_delay, missing_item, damaged_item, change_address, … (+18 more)

support_topic — Commerce support topic (commerce)

topic: billing, account_access, shipping, returns, technical_outage, onboarding, integrations, security, compliance, pricing, … (+2 more)

document_type — Document type (records)

doc_type: invoice, receipt, contract, resume, bank_statement, clinical_note, syllabus, news_article, lease, research_abstract, … (+2 more)

review_sentiment — Review sentiment (reviews)

sentiment: negative, neutral, positive

agent_handoff — Assistant handoff (assistants)

should_handoff: yes, no

email_triage — Email triage (workplace)

category: billing, support, sales, legal, recruiting, calendar, newsletter, security
action: reply, approve, fyi, escalate
needs_reply: yes, no
is_phishing: yes, no

ticket_route — Commerce ticket route (commerce)

queue: billing_disputes, refunds, shipping, returns, identity, payments, outages, mobile_app, api, account_closure, … (+6 more)
urgency: low, normal, high, critical
contains_pii: yes, no

product_feedback — Product feedback (commerce)

feedback_type: bug, feature_request, how_to, praise
product_area (multi-label): billing, mobile, search, checkout, shipping, notifications, admin, integrations

banking_intent — Banking intent (finance)

intent: balance, transfer, card_lost, dispute_charge, loan_payment, mortgage, wire, statement, fraud_alert, branch_appointment, … (+6 more)

clinic_request — Clinic request (health)

request: appointment, refill, lab_results, referral, billing, insurance, records, symptom_question, cancel_visit, provider_message, … (+1 more)
urgent: yes, no

travel_request — Travel request (travel)

intent: new_booking, change_flight, cancel_trip, seat, baggage, refund, hotel, car, visa_docs, loyalty, … (+2 more)

news_topic — News topic (news)

topic: politics, business, technology, science, health, sports, entertainment, world, culture, weather

paper_field — Paper field (science)

field: biology, medicine, chemistry, physics, computer_science, mathematics, earth_science, engineering, economics, other

sports_recap — Sports recap (sports)

sport: football, basketball, baseball, soccer, tennis, cricket, motorsport, other
result: win, loss, draw, upcoming
upset: yes, no

restaurant_review — Restaurant review (food)

sentiment: negative, neutral, positive
aspects (multi-label): food, service, price, ambiance, wait, cleanliness

benefits_request — Benefits request (civic)

program: unemployment, health_coverage, food_assistance, housing_aid, childcare, pension, disability, tax_credit, other
asking_status: yes, no

screen_tags — Screen tags (entertainment)

genres (multi-label): comedy, drama, action, documentary, horror, romance, sci_fi, kids
format: film, series

Examples

One worked example per domain (truncated for readability where long).

Commerce support intent (`support_intent`)

Route a live message before a human sees it. Refunds, cancellations, and shipping delays share one inbox and one label set.

[channel] email
[utterance] Dear support team, I reached out last week regarding issues with my monthly subscription renewal for order SUB-204859, and after following the suggested troubleshooting steps, the problem seems unresolved. The latest invoice dated 2024-04-15 still shows a charge of ¥5,400 including consumption tax, which I did not authorize due to the service interruption. Before I request a refund, could …

Gold labels:

intent → refund_request

Commerce support topic (`support_topic`)

The same inbox, a different question: which product area owns the failure.

[subject] Product page sign-in error during renewal review
[body] During our latest review, Meera at Kaveri Homeware found the product browsing screen failing before any order was placed. The merchant account showed an account-lockout message, although the failed sign-in flow began after upgrading the Checkout SDK from v4.8 to v4.9; /catalog/session and /customer/token now return 502. Subscription renewal for the …

Gold labels:

topic → technical_outage

Document type (`document_type`)

Inboxes mix invoices, contracts, and resumes. The type is the gate in front of extraction.

[document]
ARCHIVED CAREER PROFILE — RIVERGLEN COMMUNITY LEARNING HUB
Record ref: Q7N-204 | Filed: 14 Aug 2023 | Prepared by: L. Maren, program coordinator

For the instructor’s review, this historical profile summarizes Elena Voss’s service as a literacy tutor and workshop assistant. She planned beginner reading sessions, explained assignments in clear language, and kept attendance totals for evening groups. From …

Gold labels:

doc_type → resume

Review sentiment (`review_sentiment`)

A product can be praised and rejected in one paragraph. The label is what a reply policy reads.

[course] We just finished our second run of the Little Explorers Science Lab with our kids after loving the trial class last season. Was the menu of hands-on demos and take-home snack kits as good as advertised? Not really. Food quality: The promised "healthy fruit platters" were mostly store-bought granola bars and juice boxes this time, so that was a letdown. Service at the tables: The helpers answered questions …

Gold labels:

sentiment → neutral

Assistant handoff (`agent_handoff`)

Most turns stay automated. A repeated complaint or a request the policy forbids should leave the flow.

[policy] The Meadowbrook Clinic appointment assistant may assist with booking, changing, or confirming routine consultations and provide documents about standard pre-visit requirements. It must escalate requests involving medical emergencies, medication refills, billing corrections, or interactions that require a licensed clinician or office manager.
[user] I've got a yearly checkup scheduled at the Parkside branch …

Gold labels:

should_handoff → no

Email triage (`email_triage`)

One message, four decisions: what it is, what to do, whether anyone must reply, and whether it is a lure.

From: Lydia Harper <lydia.harper@cogentware.example>
To: admin@impulselectronics.example
Subject: Coordination—rescheduling quarterly review deadline

Hello Admin Team,

Summary: The project review that was previously set for today must be moved to accommodate a late-breaking platform update. We request confirmation for a new time slot before 3 p.m. Eastern today so stakeholders can be notified on schedule. …

Gold labels:

category → calendar
action → escalate
needs_reply → yes
is_phishing → no

Commerce ticket route (`ticket_route`)

The customer describes a problem, not a department. Queue, urgency, and whether the text contains personal data are scored together.

[channel] chat
[subject] Follow-up on compliance check for delayed invoice payment
[body] After the latest order update, I noticed that the payment for invoice number INV-209384 issued on 2027-03-15 at 13:45 UTC has not been processed due to account verification issues. Despite submitting all required documentation previously, the system still restricts order modifications and payment submissions. The message …

Gold labels:

queue → compliance
urgency → low
contains_pii → yes

Product feedback (`product_feedback`)

Bug, feature, or praise, plus the product areas that apply. product_area is multi-label.

[source] app review
[text] I was trying to finalize my purchase during the holiday rush using my tablet’s browser, but the notifications about my multiple shipment orders kept repeating and were out of sync with the payment authorization updates. It made tracking which items had their subscriptions renewed a mess. Each notification seemed to loop, forcing me to cross-check everything manually before hitting submit …

Gold labels:

feedback_type → bug
product_area → notifications

Banking intent (`banking_intent`)

A branch visit can mix a wire, a loan, and a hold. The label is the operation the core system should open.

[channel] branch
[utterance] I’m here at the Riverside branch regarding the auto loan my new employer, Ferris & Riggs Group, recently approved and started funding through my Harborline business banking profile. I authorized an international wire transfer for the down payment to AutoBazaar GmbH in Germany last Friday, and I see the funds are held under an authorization but have not yet posted as a completed …

Gold labels:

intent → other

Clinic request (`clinic_request`)

Patients describe a symptom and the thing they want in the same sentence. The desk needs the request and whether it is urgent.

[channel] chat
[message] When I first spoke with the clinic last month, the nurse told me that Dr. Hammond had submitted the referral for my post-op physical therapy directly to the rehab center, but I haven't received any scheduling info yet. I had a follow-up appointment planned for next week, but I missed it because of a work emergency. I know I should be doing some light exercises related to my amoxicillin …

Gold labels:

request → other
urgent → yes

Travel request (`travel_request`)

A change, a seat, and a hotel voucher look alike in chat and trigger different inventory calls.

[channel] app
[trip] QL549, terminal 2, Impora to Zavidin via Altor, departing 8 July
[message] When considering my upgraded premium seat on QL549 next week, I realized I might need assistance with accommodations due to the overnight layover in Altor — could you please confirm if a hotel voucher is included with my fare class? Also, I've already submitted my visa documents for this trip, but apologies if any details …

Gold labels:

intent → hotel

News topic (`news_topic`)

A headline has to land in a section before it is ranked. The label set is the section list.

[headline] Ferry services resume as rural Eastport hosts annual "Wind Songs" festival
[lead] A service disruption left thousands stranded along the Eastport coast after mechanical issues forced ferry cancellations during last week's storm. Organizers now say the local "Wind Songs" music festival, scheduled for 5 April, will proceed as vessels return—responding to community calls to boost visitor numbers following a …

Gold labels:

topic → world

Paper field (`paper_field`)

Title and abstract in, the field the library or the reviewer queue should file.

[title] A Microfluidic Reactor for Confidence-Calibrated Enzyme Cascade Synthesis
[abstract] Conventional batch synthesis of the target fluorinated ester requires prolonged stirring and gives variable conversion under laboratory conditions. We designed a segmented-flow microreactor with immobilized lipase and an inline optical detector, then compared it with the established stirred-vessel protocol using matched …

Gold labels:

field → chemistry

Sports recap (`sports_recap`)

Sport, result, and whether the result was an upset, from one recap.

[recap]
At the summit of the Crystalline Peaks Circuit, the Zenith Riders and Aurora Speedsters locked horns in a championship showdown that ended in a thrilling stalemate. The high-altitude track was bone-dry, offering perfect grip, yet the relentless pace challenged tire wear across the field. Early series saw frequent lead swaps, with the Riders leveraging aggressive overtakes, while the Speedsters countered with …

Gold labels:

sport → motorsport
result → draw
upset → yes

Restaurant review (`restaurant_review`)

Sentiment is one label. Aspects are several: food, service, price, wait. aspects is multi-label.

[place] Trattoria Lucca, Orvieto Crossing
[text] Compared with my last visit, our drive-through stop at Trattoria Lucca before the jazz concert was disappointing. Crisp garlic hit first, but the handmade tagliatelle was overcooked, missing that al dente bite. My serving arrived tepid, and the basil was wilted. Buon appetito felt forced this time.

Gold labels:

sentiment → negative
aspects → food

Benefits request (`benefits_request`)

Which program the person is asking about, and whether they are asking for the status of a claim already filed.

[office] Ravenwood County Benefits Department
[message] My existing unemployment claim, case number UB-55902, is pending review since I submitted all required documents last month. I provided a medical certification from Pinewood Clinic, as I had a temporary disability that delayed my job search. I recently relocated to a new apartment, so I’m uncertain if my updated address is properly recorded. I am also preparing …

Gold labels:

program → unemployment
asking_status → yes

Screen tags (`screen_tags`)

A title needs a format and every genre that fits. genres is multi-label.

At the bustling heart of Viridian City in 1912, Project LARK-27 bridges neighborhoods through a public library's curious new reading circles. Each episode intertwines the daily adventures of Sofina, a young nanny, and Meena, the spirited child she cares for, as they encounter a parade of inventive library contests, rooftop rain gardens, and trams packed with storytellers. Drawing from community traditions and …

Gold labels:

genres → kids, drama
format → series

Benchmark

Exact-match accuracy on the held-out test split: 17 domains, 300 examples each, the same text and the same candidate labels for every model. GLiNER2.5-Decide is the English model. GLiNER2.5-multi-Decide is the multilingual model, scored here on English.

Model	Avg
GLiNER2.5-Decide	60.2%
GLiNER2 XL (1B)	59.6%
JevK5	57.6%
GLiNER2.5-multi-Decide	56.7%
SemIf (Qwen3.5-4B)	56.4%
GLiFormer large-v1	49.0%
Laya Router	46.6%

Split	Examples per domain	Where
Development	100	This dataset
Test	300	Held out

Loading

One domain:

from datasets import load_dataset

ds = load_dataset("fastino/fast-decisions", "support_intent", split="train")

Every domain:

from datasets import load_dataset, concatenate_datasets

domains = [
    "support_intent", "support_topic", "document_type", "review_sentiment",
    "agent_handoff", "email_triage", "ticket_route", "product_feedback",
    "banking_intent", "clinic_request", "travel_request", "news_topic",
    "paper_field", "sports_recap", "restaurant_review", "benefits_request",
    "screen_tags",
]
all_ds = concatenate_datasets([
    load_dataset("fastino/fast-decisions", name, split="train") for name in domains
])

Exact match against GLiNER2.5-Decide. A single-label head returns a string. A multi-label head returns a list. Both are compared as sets.

from gliner2 import AutoExtractor

model = AutoExtractor.from_pretrained("fastino/GLiNER2.5-Decide")

hits = total = 0
for row in ds:
    for head in row["output"]["classifications"]:
        pred = model.classify_text(row["input"], {head["task"]: head["labels"]})
        got = pred[head["task"]]
        if isinstance(got, str):
            got = [got]
        elif isinstance(got, dict):
            got = [got["label"]]
        hits += sorted(got) == sorted(head["true_label"])
        total += 1

For text that is not English, load GLiNER2.5-multi-Decide the same way.

Details

What it is: seventeen operational classification domains for GLiNER2.5-Decide
What it is not: a reasoning exam or a copy of a public benchmark
Published split: development, 100 rows per domain
Scored split: test, 300 rows per domain, held out
License: Apache 2.0

Names, account numbers, and amounts in the text are fictional.

Citation

@misc{zaratiana2025gliner2efficientmultitaskinformation,
      title={GLiNER2: An Efficient Multi-Task Information Extraction System with Schema-Driven Interface},
      author={Urchade Zaratiana and Gil Pasternak and Oliver Boyd and George Hurn-Maloney and Ash Lewis},
      year={2025},
      eprint={2507.18546},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2507.18546},
}

classification

decision-model

gliner2

multi-task

schema-driven

Contributors

urchade

5 commits

Similar repos

fastino/GLiNER2.5-Decide

Model

GLiNER2.5-Decide

184

clinc/clinc_oos

Dataset

Dataset Card for CLINC150

avbiswas/bev-decision-150K

Dataset

BEV 150K Decision Mix

fastino/gliner2.5-multi-v1

Model

GLiNER2.5 Multi: Unified Schema-Based Information Extraction

258

fastino/gliner2.5-base-v1

Model

GLiNER2.5 Base: Unified Schema-Based Information Extraction

fastino/gliner2.5-small-v1

Model

GLiNER2.5 Small: Unified Schema-Based Information Extraction

PolyAI/banking77

Dataset

Dataset Card for BANKING77

100

Mapika/decider-2b

Model

decider-2b: typed decisions with calibrated probabilities in one forward pass

fastino/fast-decisions

Dataset

Fast Decisions

5 commits

updated Sep 24, 2026

See the code

README

Fast Decisions

Schema

Every row has this shape:

{
  "input": "<free text>",
  "output": {
    "classifications": [
      {
        "task": "<head name>",
        "true_label": [
          "<gold label(s)>"
        ],
        "labels": [
          "<label>",
          "..."
        ],
        "multi_label": false
      }
    ]
  }
}

true_label is always a list of strings, even for single-label heads.
labels is the full candidate set for that head — pass it to your model as-is; do not shuffle out the gold label.
Rows are classification only. There is no entity field.
Ordinal heads (0–5, 0–10) use string digits, not integers.

Domains

File	Domain	Group	Heads	Multi-label heads	Rows
`support_intent.jsonl`	Commerce support intent	commerce	1	—	100
`support_topic.jsonl`	Commerce support topic	commerce	1	—	100
`document_type.jsonl`	Document type	records	1	—	100
`review_sentiment.jsonl`	Review sentiment	reviews	1	—	100
`agent_handoff.jsonl`	Assistant handoff	assistants	1	—	100
`email_triage.jsonl`	Email triage	workplace	4	—	100
`ticket_route.jsonl`	Commerce ticket route	commerce	3	—	100
`product_feedback.jsonl`	Product feedback	commerce	2	`product_area`	100
`banking_intent.jsonl`	Banking intent	finance	1	—	100
`clinic_request.jsonl`	Clinic request	health	2	—	100
`travel_request.jsonl`	Travel request	travel	1	—	100
`news_topic.jsonl`	News topic	news	1	—	100
`paper_field.jsonl`	Paper field	science	1	—	100
`sports_recap.jsonl`	Sports recap	sports	3	—	100
`restaurant_review.jsonl`	Restaurant review	food	2	`aspects`	100
`benefits_request.jsonl`	Benefits request	civic	2	—	100
`screen_tags.jsonl`	Screen tags	entertainment	2	`genres`	100

Total: 1,700 rows across 17 files, 100 rows each.

Task heads and label sets

Full label inventory per head. Use these exact strings as the candidate set at inference time.

support_intent — Commerce support intent (commerce)

intent: order_status, refund_request, cancel_subscription, update_payment, login_problem, password_reset, shipping_delay, missing_item, damaged_item, change_address, … (+18 more)

support_topic — Commerce support topic (commerce)

topic: billing, account_access, shipping, returns, technical_outage, onboarding, integrations, security, compliance, pricing, … (+2 more)

document_type — Document type (records)

doc_type: invoice, receipt, contract, resume, bank_statement, clinical_note, syllabus, news_article, lease, research_abstract, … (+2 more)

review_sentiment — Review sentiment (reviews)

sentiment: negative, neutral, positive

agent_handoff — Assistant handoff (assistants)

should_handoff: yes, no

email_triage — Email triage (workplace)

category: billing, support, sales, legal, recruiting, calendar, newsletter, security
action: reply, approve, fyi, escalate
needs_reply: yes, no
is_phishing: yes, no

ticket_route — Commerce ticket route (commerce)

queue: billing_disputes, refunds, shipping, returns, identity, payments, outages, mobile_app, api, account_closure, … (+6 more)
urgency: low, normal, high, critical
contains_pii: yes, no

product_feedback — Product feedback (commerce)

feedback_type: bug, feature_request, how_to, praise
product_area (multi-label): billing, mobile, search, checkout, shipping, notifications, admin, integrations

banking_intent — Banking intent (finance)

intent: balance, transfer, card_lost, dispute_charge, loan_payment, mortgage, wire, statement, fraud_alert, branch_appointment, … (+6 more)

clinic_request — Clinic request (health)

request: appointment, refill, lab_results, referral, billing, insurance, records, symptom_question, cancel_visit, provider_message, … (+1 more)
urgent: yes, no

travel_request — Travel request (travel)

intent: new_booking, change_flight, cancel_trip, seat, baggage, refund, hotel, car, visa_docs, loyalty, … (+2 more)

news_topic — News topic (news)

topic: politics, business, technology, science, health, sports, entertainment, world, culture, weather

paper_field — Paper field (science)

field: biology, medicine, chemistry, physics, computer_science, mathematics, earth_science, engineering, economics, other

sports_recap — Sports recap (sports)

sport: football, basketball, baseball, soccer, tennis, cricket, motorsport, other
result: win, loss, draw, upcoming
upset: yes, no

restaurant_review — Restaurant review (food)

sentiment: negative, neutral, positive
aspects (multi-label): food, service, price, ambiance, wait, cleanliness

benefits_request — Benefits request (civic)

program: unemployment, health_coverage, food_assistance, housing_aid, childcare, pension, disability, tax_credit, other
asking_status: yes, no

screen_tags — Screen tags (entertainment)

genres (multi-label): comedy, drama, action, documentary, horror, romance, sci_fi, kids
format: film, series

Examples

One worked example per domain (truncated for readability where long).

Commerce support intent (`support_intent`)

Route a live message before a human sees it. Refunds, cancellations, and shipping delays share one inbox and one label set.

[channel] email
[utterance] Dear support team, I reached out last week regarding issues with my monthly subscription renewal for order SUB-204859, and after following the suggested troubleshooting steps, the problem seems unresolved. The latest invoice dated 2024-04-15 still shows a charge of ¥5,400 including consumption tax, which I did not authorize due to the service interruption. Before I request a refund, could …

Gold labels:

intent → refund_request

Commerce support topic (`support_topic`)

The same inbox, a different question: which product area owns the failure.

[subject] Product page sign-in error during renewal review
[body] During our latest review, Meera at Kaveri Homeware found the product browsing screen failing before any order was placed. The merchant account showed an account-lockout message, although the failed sign-in flow began after upgrading the Checkout SDK from v4.8 to v4.9; /catalog/session and /customer/token now return 502. Subscription renewal for the …

Gold labels:

topic → technical_outage

Document type (`document_type`)

Inboxes mix invoices, contracts, and resumes. The type is the gate in front of extraction.

[document]
ARCHIVED CAREER PROFILE — RIVERGLEN COMMUNITY LEARNING HUB
Record ref: Q7N-204 | Filed: 14 Aug 2023 | Prepared by: L. Maren, program coordinator

For the instructor’s review, this historical profile summarizes Elena Voss’s service as a literacy tutor and workshop assistant. She planned beginner reading sessions, explained assignments in clear language, and kept attendance totals for evening groups. From …

Gold labels:

doc_type → resume

Review sentiment (`review_sentiment`)

A product can be praised and rejected in one paragraph. The label is what a reply policy reads.

[course] We just finished our second run of the Little Explorers Science Lab with our kids after loving the trial class last season. Was the menu of hands-on demos and take-home snack kits as good as advertised? Not really. Food quality: The promised "healthy fruit platters" were mostly store-bought granola bars and juice boxes this time, so that was a letdown. Service at the tables: The helpers answered questions …

Gold labels:

sentiment → neutral

Assistant handoff (`agent_handoff`)

Most turns stay automated. A repeated complaint or a request the policy forbids should leave the flow.

[policy] The Meadowbrook Clinic appointment assistant may assist with booking, changing, or confirming routine consultations and provide documents about standard pre-visit requirements. It must escalate requests involving medical emergencies, medication refills, billing corrections, or interactions that require a licensed clinician or office manager.
[user] I've got a yearly checkup scheduled at the Parkside branch …

Gold labels:

should_handoff → no

Email triage (`email_triage`)

One message, four decisions: what it is, what to do, whether anyone must reply, and whether it is a lure.

From: Lydia Harper <lydia.harper@cogentware.example>
To: admin@impulselectronics.example
Subject: Coordination—rescheduling quarterly review deadline

Hello Admin Team,

Summary: The project review that was previously set for today must be moved to accommodate a late-breaking platform update. We request confirmation for a new time slot before 3 p.m. Eastern today so stakeholders can be notified on schedule. …

Gold labels:

category → calendar
action → escalate
needs_reply → yes
is_phishing → no

Commerce ticket route (`ticket_route`)

The customer describes a problem, not a department. Queue, urgency, and whether the text contains personal data are scored together.

[channel] chat
[subject] Follow-up on compliance check for delayed invoice payment
[body] After the latest order update, I noticed that the payment for invoice number INV-209384 issued on 2027-03-15 at 13:45 UTC has not been processed due to account verification issues. Despite submitting all required documentation previously, the system still restricts order modifications and payment submissions. The message …

Gold labels:

queue → compliance
urgency → low
contains_pii → yes

Product feedback (`product_feedback`)

Bug, feature, or praise, plus the product areas that apply. product_area is multi-label.

[source] app review
[text] I was trying to finalize my purchase during the holiday rush using my tablet’s browser, but the notifications about my multiple shipment orders kept repeating and were out of sync with the payment authorization updates. It made tracking which items had their subscriptions renewed a mess. Each notification seemed to loop, forcing me to cross-check everything manually before hitting submit …

Gold labels:

feedback_type → bug
product_area → notifications

Banking intent (`banking_intent`)

A branch visit can mix a wire, a loan, and a hold. The label is the operation the core system should open.

[channel] branch
[utterance] I’m here at the Riverside branch regarding the auto loan my new employer, Ferris & Riggs Group, recently approved and started funding through my Harborline business banking profile. I authorized an international wire transfer for the down payment to AutoBazaar GmbH in Germany last Friday, and I see the funds are held under an authorization but have not yet posted as a completed …

Gold labels:

intent → other

Clinic request (`clinic_request`)

Patients describe a symptom and the thing they want in the same sentence. The desk needs the request and whether it is urgent.

[channel] chat
[message] When I first spoke with the clinic last month, the nurse told me that Dr. Hammond had submitted the referral for my post-op physical therapy directly to the rehab center, but I haven't received any scheduling info yet. I had a follow-up appointment planned for next week, but I missed it because of a work emergency. I know I should be doing some light exercises related to my amoxicillin …

Gold labels:

request → other
urgent → yes

Travel request (`travel_request`)

A change, a seat, and a hotel voucher look alike in chat and trigger different inventory calls.

[channel] app
[trip] QL549, terminal 2, Impora to Zavidin via Altor, departing 8 July
[message] When considering my upgraded premium seat on QL549 next week, I realized I might need assistance with accommodations due to the overnight layover in Altor — could you please confirm if a hotel voucher is included with my fare class? Also, I've already submitted my visa documents for this trip, but apologies if any details …

Gold labels:

intent → hotel

News topic (`news_topic`)

A headline has to land in a section before it is ranked. The label set is the section list.

[headline] Ferry services resume as rural Eastport hosts annual "Wind Songs" festival
[lead] A service disruption left thousands stranded along the Eastport coast after mechanical issues forced ferry cancellations during last week's storm. Organizers now say the local "Wind Songs" music festival, scheduled for 5 April, will proceed as vessels return—responding to community calls to boost visitor numbers following a …

Gold labels:

topic → world

Paper field (`paper_field`)

Title and abstract in, the field the library or the reviewer queue should file.

[title] A Microfluidic Reactor for Confidence-Calibrated Enzyme Cascade Synthesis
[abstract] Conventional batch synthesis of the target fluorinated ester requires prolonged stirring and gives variable conversion under laboratory conditions. We designed a segmented-flow microreactor with immobilized lipase and an inline optical detector, then compared it with the established stirred-vessel protocol using matched …

Gold labels:

field → chemistry

Sports recap (`sports_recap`)

Sport, result, and whether the result was an upset, from one recap.

[recap]
At the summit of the Crystalline Peaks Circuit, the Zenith Riders and Aurora Speedsters locked horns in a championship showdown that ended in a thrilling stalemate. The high-altitude track was bone-dry, offering perfect grip, yet the relentless pace challenged tire wear across the field. Early series saw frequent lead swaps, with the Riders leveraging aggressive overtakes, while the Speedsters countered with …

Gold labels:

sport → motorsport
result → draw
upset → yes

Restaurant review (`restaurant_review`)

Sentiment is one label. Aspects are several: food, service, price, wait. aspects is multi-label.

[place] Trattoria Lucca, Orvieto Crossing
[text] Compared with my last visit, our drive-through stop at Trattoria Lucca before the jazz concert was disappointing. Crisp garlic hit first, but the handmade tagliatelle was overcooked, missing that al dente bite. My serving arrived tepid, and the basil was wilted. Buon appetito felt forced this time.

Gold labels:

sentiment → negative
aspects → food

Benefits request (`benefits_request`)

Which program the person is asking about, and whether they are asking for the status of a claim already filed.

[office] Ravenwood County Benefits Department
[message] My existing unemployment claim, case number UB-55902, is pending review since I submitted all required documents last month. I provided a medical certification from Pinewood Clinic, as I had a temporary disability that delayed my job search. I recently relocated to a new apartment, so I’m uncertain if my updated address is properly recorded. I am also preparing …

Gold labels:

program → unemployment
asking_status → yes

Screen tags (`screen_tags`)

A title needs a format and every genre that fits. genres is multi-label.

At the bustling heart of Viridian City in 1912, Project LARK-27 bridges neighborhoods through a public library's curious new reading circles. Each episode intertwines the daily adventures of Sofina, a young nanny, and Meena, the spirited child she cares for, as they encounter a parade of inventive library contests, rooftop rain gardens, and trams packed with storytellers. Drawing from community traditions and …

Gold labels:

genres → kids, drama
format → series

Benchmark

Model	Avg
GLiNER2.5-Decide	60.2%
GLiNER2 XL (1B)	59.6%
JevK5	57.6%
GLiNER2.5-multi-Decide	56.7%
SemIf (Qwen3.5-4B)	56.4%
GLiFormer large-v1	49.0%
Laya Router	46.6%

Split	Examples per domain	Where
Development	100	This dataset
Test	300	Held out

Loading

One domain:

from datasets import load_dataset

ds = load_dataset("fastino/fast-decisions", "support_intent", split="train")

Every domain:

from datasets import load_dataset, concatenate_datasets

domains = [
    "support_intent", "support_topic", "document_type", "review_sentiment",
    "agent_handoff", "email_triage", "ticket_route", "product_feedback",
    "banking_intent", "clinic_request", "travel_request", "news_topic",
    "paper_field", "sports_recap", "restaurant_review", "benefits_request",
    "screen_tags",
]
all_ds = concatenate_datasets([
    load_dataset("fastino/fast-decisions", name, split="train") for name in domains
])

Exact match against GLiNER2.5-Decide. A single-label head returns a string. A multi-label head returns a list. Both are compared as sets.

from gliner2 import AutoExtractor

model = AutoExtractor.from_pretrained("fastino/GLiNER2.5-Decide")

hits = total = 0
for row in ds:
    for head in row["output"]["classifications"]:
        pred = model.classify_text(row["input"], {head["task"]: head["labels"]})
        got = pred[head["task"]]
        if isinstance(got, str):
            got = [got]
        elif isinstance(got, dict):
            got = [got["label"]]
        hits += sorted(got) == sorted(head["true_label"])
        total += 1

For text that is not English, load GLiNER2.5-multi-Decide the same way.

Details

What it is: seventeen operational classification domains for GLiNER2.5-Decide
What it is not: a reasoning exam or a copy of a public benchmark
Published split: development, 100 rows per domain
Scored split: test, 300 rows per domain, held out
License: Apache 2.0

Names, account numbers, and amounts in the text are fictional.

Citation

@misc{zaratiana2025gliner2efficientmultitaskinformation,
      title={GLiNER2: An Efficient Multi-Task Information Extraction System with Schema-Driven Interface},
      author={Urchade Zaratiana and Gil Pasternak and Oliver Boyd and George Hurn-Maloney and Ash Lewis},
      year={2025},
      eprint={2507.18546},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2507.18546},
}

classification

decision-model

gliner2

multi-task

schema-driven

Similar repos

fastino/GLiNER2.5-Decide

Model

GLiNER2.5-Decide

184

clinc/clinc_oos

Dataset

Dataset Card for CLINC150

avbiswas/bev-decision-150K

Dataset

BEV 150K Decision Mix

fastino/gliner2.5-multi-v1

Model

GLiNER2.5 Multi: Unified Schema-Based Information Extraction

258

fastino/gliner2.5-base-v1

Model

GLiNER2.5 Base: Unified Schema-Based Information Extraction

fastino/gliner2.5-small-v1

Model

GLiNER2.5 Small: Unified Schema-Based Information Extraction

PolyAI/banking77

Dataset

Dataset Card for BANKING77

100

Mapika/decider-2b

Model

decider-2b: typed decisions with calibrated probabilities in one forward pass

Contributors

urchade

5 commits

fastino/fast-decisions

README

Fast Decisions

Schema

Domains

Task heads and label sets

Examples

Commerce support intent (support_intent)

Commerce support topic (support_topic)

Document type (document_type)

Review sentiment (review_sentiment)

Assistant handoff (agent_handoff)

Email triage (email_triage)

Commerce ticket route (ticket_route)

Product feedback (product_feedback)

Banking intent (banking_intent)

Clinic request (clinic_request)

Travel request (travel_request)

News topic (news_topic)

Paper field (paper_field)

Sports recap (sports_recap)

Restaurant review (restaurant_review)

Benefits request (benefits_request)

Screen tags (screen_tags)

Benchmark

Loading

Details

Citation

Contributors

Similar repos

fastino/fast-decisions

README

Fast Decisions

Schema

Domains

Task heads and label sets

Examples

Commerce support intent (support_intent)

Commerce support topic (support_topic)

Document type (document_type)

Review sentiment (review_sentiment)

Assistant handoff (agent_handoff)

Email triage (email_triage)

Commerce ticket route (ticket_route)

Product feedback (product_feedback)

Banking intent (banking_intent)

Clinic request (clinic_request)

Travel request (travel_request)

News topic (news_topic)

Paper field (paper_field)

Sports recap (sports_recap)

Restaurant review (restaurant_review)

Benefits request (benefits_request)

Screen tags (screen_tags)

Benchmark

Loading

Details

Citation

Similar repos

Contributors

Commerce support intent (`support_intent`)

Commerce support topic (`support_topic`)

Document type (`document_type`)

Review sentiment (`review_sentiment`)

Assistant handoff (`agent_handoff`)

Email triage (`email_triage`)

Commerce ticket route (`ticket_route`)

Product feedback (`product_feedback`)

Banking intent (`banking_intent`)

Clinic request (`clinic_request`)

Travel request (`travel_request`)

News topic (`news_topic`)

Paper field (`paper_field`)

Sports recap (`sports_recap`)

Restaurant review (`restaurant_review`)

Benefits request (`benefits_request`)

Screen tags (`screen_tags`)

Commerce support intent (`support_intent`)

Commerce support topic (`support_topic`)

Document type (`document_type`)

Review sentiment (`review_sentiment`)

Assistant handoff (`agent_handoff`)

Email triage (`email_triage`)

Commerce ticket route (`ticket_route`)

Product feedback (`product_feedback`)

Banking intent (`banking_intent`)

Clinic request (`clinic_request`)

Travel request (`travel_request`)

News topic (`news_topic`)

Paper field (`paper_field`)

Sports recap (`sports_recap`)

Restaurant review (`restaurant_review`)

Benefits request (`benefits_request`)

Screen tags (`screen_tags`)