Datasets:
Access to TUT-100 requires Tuteliq AB approval
TUT-100 is Tuteliq AB's confidential, proprietary, held-out evaluation dataset. The card is public; the data is gated. Access is request-only and granted solely at Tuteliq's discretion after manual review, governed by the TUT-100 Dataset Access Agreement v1.0 (reproduced in full on this card). Submitting a request does not guarantee access.
Access to TUT-100 is governed by the TUT-100 Dataset Access Agreement v1.0 (Tuteliq AB, org. no. 559481-1217), reproduced in full in the dataset card below. By requesting access you, on behalf of yourself and your organization, agree to be bound by that Agreement, and in particular:
- Confidentiality - the Dataset is Tuteliq's confidential, proprietary information; you will not disclose, publish, redistribute, mirror, deposit in any public repository, or share it or any portion of it, in any form (Agreement s.2).
- Permitted use only - evaluation, benchmarking, and vendor comparison of child-safety detection systems, for the specific purpose approved in writing (s.3).
- No training - you will NOT train, fine-tune, distil, or develop any model on the Dataset, and will prevent any leakage into a training corpus; it is a held-out evaluation set (s.4.1).
- Security - encrypt at rest (AES-256 or equivalent), restrict to named Authorized Users (max 5), MFA + full-disk encryption, no unencrypted cloud sync, keep access logs (s.5).
- Breach - report any actual or suspected breach within 24 hours to legal@tuteliq.ai (s.6).
- Attribution & IP - Tuteliq retains all IP; cite as required; publishing case contents requires prior written approval (s.7).
- Term - 12 months, no automatic renewal, annual renewal required 60 days before expiry; delete all copies on termination with written certification (s.9-s.10).
- Governing law - Swedish law; Swedish courts (s.12).
Approved recipients receive gated access to this Hugging Face repository, bound to the approved account and Authorized Users list, and must execute the full Agreement. Access is request-only and granted solely at Tuteliq AB's discretion after manual review. Providing false information voids any grant.
Log in or Sign Up to review the conditions and access this dataset content.
- ⚠️ Held-out — do NOT train on this data
- Motivation & design
- Tactic taxonomy (7 tactics)
- Scoring
- Contents
- Provenance, ethics & safety
- Intended use / out of scope
- Access
- Citation
- Contact
- Full Access Agreement (v1.0)
- 1. Parties and Definitions
- 2. Confidentiality and Restricted Disclosure
- 3. Permitted Uses
- 4. Prohibited Uses
- 5. Data Handling and Security
- 6. Breach Notification
- 7. Intellectual Property and Attribution
- 8. Access Request and Registration
- 9. Term and Renewal
- 10. Termination
- 11. Liability Limitation
- 12. Governing Law and Dispute Resolution
- 13. Compliance and Audit
- 14. Amendments and Notices
- 15. Entire Agreement
- 16. Severability
- 17. Signatures
- Appendix A: Authorized Users
- Appendix B: Data Handling Attestation
TUT-100 — Tuteliq Grooming-Detection Benchmark
Maintained by Tuteliq AB. Confidential, held-out evaluation dataset. The card is public; the data is gated (manual approval). All rights reserved — do not redistribute.
TUT-100 is Tuteliq's benchmark for online child-grooming detection. It is designed to measure trajectory-aware detection — the recognition of grooming as it unfolds across a conversation — and to expose the gap between single-message content classifiers and detectors that reason over the full interaction. It is the held-out evaluation set for the tuteliq/tuteliq-detect-24b-v6 model family and for cross-vendor Trust & Safety comparison.
- Maintainer / author: Tuteliq AB
- Version: v1.0 (git tag
v1.0, release 2026-05-29) - Domain: online child grooming (text conversations)
- Languages: 16 (English + 15 EU languages)
- Content: 100% synthetic; no real PII
⚠️ Held-out — do NOT train on this data
This set exists to measure detectors. Training on it, directly or via leakage, invalidates every result reported against it. Keep it out of all training, fine-tuning, and pre-training corpora, and leak-guard training data against it (e.g. an 8-word shingle check).
Motivation & design
Single-message moderation classifiers score each message in isolation and systematically miss grooming, whose harm emerges from an escalating trajectory (rapport → flattery → gift-giving → secrecy → isolation → boundary-pushing → photo/meeting requests) rather than any one message. TUT-100 is built to reward trajectory-aware detection and to make the single-message gap measurable: its flagship item is a 100-turn escalation with per-turn tactic anchors and a documented threshold-crossing point, supported by shorter positive, adversarial, and hard-negative sets.
Tactic taxonomy (7 tactics)
Per-turn labels use Tuteliq's grooming tactic taxonomy:
| Tactic | Definition |
|---|---|
flattery |
Praise with no ask attached — singling the child out as special/talented/above peers. |
gift_giving |
Offering gifts, money, or in-game items to build obligation. |
secrecy_request |
Pushing to conceal the contact/gift/meeting/relationship from parents, carers, or teachers (incl. platform-switching framed as hiding). |
isolation |
Denigrating the child's real friends/family/trusted adults to make the child depend on the adult. |
boundary_pushing |
Testing or eroding a stated limit — repeating after a refusal, escalating after a small yes, normalising rule-breaking. |
photo_request |
Soliciting images of the child. |
meeting_request |
Attempting to arrange an in-person or off-platform meeting. |
Scoring
- Conversation-level predatory verdict (risk-score thresholded): precision / recall / F1 and AUC-ROC.
- Per-tactic F1 across the 7 tactics, from the model's per-message analysis, scored against the per-turn gold anchors on adult turns (macro over tactics).
- Threshold-crossing turn and critical-lock turn for latency-of-detection analysis.
Reference harness: runner.py / scorer.py and the per-tactic scorer (tut100/score_tactics_f1.py) in the Tuteliq packages/api/benchmark/ repository.
Contents
| Path | Cases | Label | Description |
|---|---|---|---|
TUT-100-V1_benchmark.json |
1 | grooming_positive | Flagship 100-turn synthetic predatory-grooming trajectory with per-turn tactic anchors + threshold-crossing ground truth |
positives/ |
15 | grooming_positive | Grooming trajectories (fast + slow escalation) |
adversarial/ |
10 | grooming_positive | Obfuscated grooming (euphemism, slow-burn, platform-hopping) |
benign_large/ |
60 | benign | Hard negatives (parent/coach/peer chats with surface-form look-alikes) |
multilingual/ |
10 | grooming_positive | In-language grooming across EU languages |
controls/ |
27 | mixed | Boundary/control cases for precision scoring |
eval_gate_positives/ |
35 | grooming_positive | Ship-gate positive set |
Each conversation file: {benchmark_name, label, conversation:[{turn, speaker, content, tactic_expected}], ground_truth, provenance}.
Provenance, ethics & safety
100% synthetic (synthetic_composite). Per the benchmark manifest: assembled from aggregate research patterns in academic literature, NGO casework summaries, and historical regulatory-enforcement materials — no real conversation reproduced, and no real names, locations, handles, schools, or platforms appear. The taxonomy and trajectory design reflect expert criminological review. Contains no real PII and no user content. The dataset depicts grooming tactics at an indicator level for the sole purpose of evaluating protective detection systems.
Intended use / out of scope
- In scope: evaluating and comparing grooming-detection systems under Tuteliq approval.
- Out of scope: any model training/fine-tuning; any operational or non-evaluation use; redistribution.
Access
The dataset card is public; the data is gated with manual approval. Request access via the button above and complete the agreement. Every request is reviewed by Tuteliq AB; no data is downloadable without an approved grant.
Citation
@misc{tuteliq_tut100_2026,
title = {TUT-100: A Trajectory-Aware Benchmark for Online Child-Grooming Detection},
author = {Tuteliq AB},
year = {2026},
note = {Held-out evaluation dataset, v1.0. Private/gated. Contact Tuteliq AB for access.},
url = {https://huggingface.co/datasets/tuteliq/tut-100}
}
Contact
- Web: https://tuteliq.ai
- Research & dataset-access enquiries: research@tuteliq.ai
- Legal / breach notification: legal@tuteliq.ai
Access requests are handled through the gated request flow above; general questions and research collaboration enquiries are welcome at research@tuteliq.ai.
© Tuteliq AB. Governed by the Tuteliq Partner License.
Full Access Agreement (v1.0)
The complete governing agreement. Approved recipients must execute it; the summary at the access gate is not a substitute for these terms.
TUT-100 Dataset Access Agreement
Tuteliq AB
Organization Number: 5594811217
Address: Sweden
1. Parties and Definitions
1.1 Parties
- Provider: Tuteliq AB ("Tuteliq"), a Swedish company
- Recipient: The individual or organization identified in Section 8 ("Requestor")
1.2 Definitions
"Dataset" means the TUT-100 Grooming Detection Benchmark, a held-out evaluation set consisting of:
- 1 flagship 100-turn synthetic grooming trajectory
- 15 grooming positive trajectories
- 10 adversarial/obfuscated grooming cases
- 60 hard-negative benign conversations
- 10 multilingual grooming cases in EU languages
- 27 boundary/control cases
- 35 eval-gate positive cases
- Total: 158 synthetic conversation files with per-turn annotations and ground truth labels
- All files in JSON format per benchmark manifest v1.0 (release date: May 29, 2026)
"Permitted Use" means evaluation, benchmarking, or vendor comparison of child-safety detection systems, as defined in Section 3.
"Prohibited Use" means any use of the Dataset other than Permitted Use, including but not limited to training, fine-tuning, or model development.
"Breach" means unauthorized access, use, disclosure, or loss of the Dataset.
"Institutional Affiliation" means employment at, or official engagement with, a university, research organization, child-safety vendor, or Tuteliq customer account holder.
2. Confidentiality and Restricted Disclosure
2.1 Confidential Nature
The Requestor acknowledges that:
- The Dataset is Tuteliq's confidential and proprietary information
- The Dataset is a held-out evaluation set for model validation
- Public disclosure or unauthorized sharing invalidates research integrity for all parties
- The Dataset contains no real personal information (100% synthetic), but represents sensitive child-safety scenarios
2.2 Recipient Obligations
Requestor agrees to:
- Treat the Dataset as strictly confidential
- Restrict access to named individuals on an approved "Authorized Users" list (Section 8)
- Store the Dataset securely and encrypt at rest (AES-256 or equivalent)
- Prevent unauthorized copying, downloading, or transfer
- Not republish, redistribute, or share the Dataset with third parties
- Not deposit the Dataset in public repositories (GitHub, Zenodo, HuggingFace, etc.)
- Destroy all copies upon termination of this agreement or upon written request
2.3 Exceptions to Confidentiality
Requestor may disclose the Dataset only:
- To named Authorized Users on the registered access list
- To legal counsel or compliance officers for purposes of enforcing this agreement
- When required by court order (with 30 days' notice to Tuteliq when legally permitted)
3. Permitted Uses
3.1 Evaluation and Benchmarking
Requestor may use the Dataset ONLY for:
- Evaluation of their own child-safety detection models
- Comparison against Tuteliq Detect or other vendor models
- Academic research on child-safety AI systems (with Institutional Affiliation)
- Internal benchmarking and performance metrics
- Regulatory compliance validation (e.g., demonstrating model performance for child safety)
3.2 Per-Turn Analysis
Requestor may conduct:
- Per-message tactic detection (via message_analysis in the Tuteliq packages/api/benchmark/ harness)
- Turn-by-turn ground-truth comparison
- Precision/recall/F1/AUC-ROC scoring using the official scoring harness
- Analysis of per-turn tactic labels (secrecy_request, isolation, boundary_pushing, flattery, photo_request, gift_giving, meeting_request)
3.3 Reporting
Requestor may publish findings as:
- Academic papers (with prior written approval from Tuteliq)
- Internal reports to their organization
- Regulatory submissions (where Dataset provenance is described as synthetic)
- Vendor comparison materials (anonymized, not naming specific cases from the Dataset)
Note: Publication of detailed case contents from the Dataset (full conversation transcripts, specific dialogue) requires Tuteliq's prior written approval.
4. Prohibited Uses
Requestor STRICTLY PROHIBITS:
4.1 Training and Fine-Tuning
- ❌ Training machine learning models on the Dataset
- ❌ Fine-tuning language models using Dataset conversations
- ❌ Using Dataset examples in prompt engineering or few-shot learning
- ❌ Leakage of Dataset examples into training data (direct or indirect)
- ❌ Constructing synthetic data "inspired by" or derived from Dataset cases
- ❌ Using the Dataset as an in-domain corpus for model adaptation
4.2 Redistribution
- ❌ Sharing the Dataset with external researchers, partners, or vendors
- ❌ Depositing the Dataset in public repositories or open-science platforms
- ❌ Selling or commercializing access to the Dataset
- ❌ Bundling the Dataset with products or services without Tuteliq permission
4.3 Competitive Intelligence
- ❌ Reverse-engineering Tuteliq's model design from Dataset annotations
- ❌ Using Dataset patterns to directly replicate Tuteliq Detect's architecture
- ❌ Extracting Dataset metadata for comparative marketing without attribution
5. Data Handling and Security
5.1 Secure Storage
Requestor agrees to:
- Encrypt the Dataset at rest using AES-256 or equivalent
- Store the Dataset on isolated, access-controlled systems
- Restrict network access to Dataset files (no cloud storage without encryption)
- Maintain audit logs of all Dataset access (who, when, what)
5.2 Authorized Users
Requestor shall:
- Limit access to named individuals (listed in Section 8)
- Remove access immediately when an Authorized User leaves the organization
- Conduct background checks or institutional verification for all Authorized Users (at Requestor's discretion)
- Notify Tuteliq of any additions or removals to the Authorized Users list within 5 business days
5.3 Workstation Requirements
The Dataset must be accessed only on:
- Workstations with full-disk encryption enabled
- Systems with multi-factor authentication for login
- Networks with firewall protection and intrusion detection
- Devices not connected to unsecured networks or public WiFi during Dataset access
5.4 No Cloud Synchronization
Requestor shall NOT:
- Sync Dataset files to cloud storage (Google Drive, Dropbox, OneDrive, iCloud, etc.)
- Upload Dataset to shared repositories or file-sharing platforms
- Store unencrypted Dataset copies on cloud servers
6. Breach Notification
6.1 Incident Reporting
If Requestor becomes aware of any actual or suspected Breach, Requestor shall:
- Notify Tuteliq immediately (within 24 hours) at legal@tuteliq.ai
- Provide detailed description of the incident (what, when, who was affected, how discovered)
- Include estimated scope (how many records, which Authorized Users involved)
- Detail containment measures already taken
6.2 Investigation and Cooperation
Requestor shall:
- Preserve all evidence related to the Breach
- Cooperate with Tuteliq's investigation into the cause and extent
- Provide forensic logs and access records upon request
- Implement remedial measures recommended by Tuteliq within 30 days
6.3 Termination Right
Breach of data security terms may result in immediate termination of this agreement and revocation of Dataset access.
7. Intellectual Property and Attribution
7.1 Ownership
Tuteliq retains all intellectual property rights in:
- The Dataset itself and its contents
- The benchmark structure, annotations, and ground truth labels
- The per-turn tactic taxonomy and labeling scheme
- The official scoring harness and evaluation methodology
7.2 Permitted Attribution
Requestor may cite the Dataset in publications as:
"We evaluated our model on the TUT-100 Grooming Detection Benchmark, a held-out evaluation set developed by Tuteliq AB. TUT-100 consists of 158 synthetic conversations with per-turn grooming tactic annotations."
Requestor shall include:
- Proper citation to Tuteliq AB as the Dataset provider
- Statement that the Dataset is held-out and not for training
- Provenance description (synthetic, no real PII)
- Benchmark version (v1.0)
8. Access Request and Registration
8.1 Requestor Information
To request access, Requestor must provide:
- Full legal name and title
- Organization name and type (academic, vendor, enterprise customer, etc.)
- Institutional affiliation verification (university email, business registration, customer account)
- Primary email address and phone number
- Brief description of Permitted Use (e.g., "benchmarking our grooming detection model")
8.2 Authorized Users List
Requestor must designate:
- Primary contact (person responsible for compliance)
- Authorized Users (maximum 5 persons per organization; higher numbers require justification)
- For each Authorized User: name, email, title, role in project
8.3 Approval Process
Tuteliq shall:
- Review access request within 10 business days
- Request additional information if needed
- Approve or deny access in writing
- Issue signed agreement and Dataset access credentials if approved
Note: Tuteliq reserves the right to deny access without providing detailed reasoning.
8.4 Access Credentials
Upon approval, Requestor receives:
- Gated access to the official Hugging Face repository (https://huggingface.co/datasets/tuteliq/tut-100), granted to the approved account
- Benchmark documentation and scoring harness reference
- Configuration file for the official evaluation harness
Access is bound to the approved Hugging Face account and the registered Authorized Users list. The Dataset must not be re-hosted, mirrored, copied out of, or redistributed beyond that gated repository.
9. Term and Renewal
9.1 Initial Term
This agreement is effective upon signature and valid for 12 months from the date of approval.
9.2 Annual Renewal
Requestor must renew access annually by submitting:
- Confirmation of continued compliance with this agreement
- Updated Authorized Users list
- Brief summary of uses conducted during the past year (publications, research outcomes, if any)
- Attestation that no Breaches occurred
- Renewed commitment to data security requirements
Renewal requests must be submitted 60 days before expiration.
9.3 Renewal Decision
Tuteliq shall:
- Review renewal requests within 20 business days
- Approve or deny renewal in writing
- Revoke access if renewal is denied or if Requestor fails to renew
Important: No automatic renewal. Access expires 12 months from approval date.
10. Termination
10.1 Termination by Tuteliq
Tuteliq may terminate this agreement immediately if:
- Requestor materially breaches confidentiality or security terms
- Unauthorized access, redistribution, or training on the Dataset is discovered
- Requestor fails to renew on schedule or fails to respond to renewal notices
- Authorized Users violate access restrictions
- Regulatory or legal obligations require termination
- Tuteliq ceases operation of the benchmark program
Tuteliq will provide written notice specifying the reason and effective date.
10.2 Termination by Requestor
Requestor may terminate this agreement with 60 days' written notice to legal@tuteliq.ai.
10.3 Effect of Termination
Upon termination, Requestor shall:
- Immediately cease all use of the Dataset
- Permanently delete all copies of the Dataset (with no recovery)
- Provide written certification of deletion within 10 days
- Cooperate with any final audit or compliance verification
- Retain confidentiality obligations indefinitely for past use
11. Liability Limitation
11.1 No Warranty
The Dataset is provided "AS-IS" without warranty of any kind. Tuteliq makes no representations regarding:
- Accuracy, completeness, or suitability of Dataset for any particular purpose
- Non-infringement of third-party rights
- Fitness for research, commercial, or regulatory use
- Performance of models trained or evaluated using the Dataset
11.2 Liability Cap
Tuteliq's total liability under this agreement shall not exceed:
- The amount paid by Requestor for access (if any), OR
- USD $0 if access is provided at no cost
Tuteliq is not liable for:
- Lost profits, revenue, or business opportunity
- Indirect, incidental, special, or consequential damages
- Third-party claims related to Requestor's use of the Dataset
- Regulatory fines or penalties arising from Requestor's Breach
12. Governing Law and Dispute Resolution
12.1 Governing Law
This agreement is governed by Swedish law, without regard to conflicts of law principles.
12.2 Venue
Any disputes arising from this agreement shall be resolved in Swedish courts. Both parties consent to the exclusive jurisdiction of Swedish courts.
12.3 Pre-Litigation Resolution
Before initiating formal proceedings, both parties agree to:
- Exchange detailed written notice of the dispute
- Attempt good-faith negotiation for 30 days
- Escalate to senior management if initial negotiation fails
13. Compliance and Audit
13.1 Audit Rights
Tuteliq reserves the right to:
- Request audit documentation of Dataset security measures
- Conduct security assessments (remote or on-site) upon 10 days' notice
- Verify Authorized Users list and access controls
- Review data handling logs and access records
Requestor shall provide requested documentation within 15 business days.
13.2 Regulatory Compliance
Requestor warrants that:
- All use of the Dataset complies with applicable laws
- Requestor has authority to enter this agreement
- No governmental or institutional approvals are required for Dataset use
13.3 Child Safety Commitment
Requestor acknowledges that the Dataset exists to improve child-safety detection systems. Requestor agrees that:
- The Dataset will be used only for legitimate child-safety research or evaluation
- Results will not be misused to harm children or undermine safety protections
- Any regulatory findings or policy changes will be shared with Tuteliq if relevant to child safety
14. Amendments and Notices
14.1 Amendments
Tuteliq may amend this agreement:
- With 30 days' written notice for material changes
- Immediately for clarifications or administrative changes
- Continued use of the Dataset constitutes acceptance of amendments
14.2 Notices
All notices shall be sent to the addresses below:
To Tuteliq:
Tuteliq AB
Legal / breach notification: legal@tuteliq.ai
Research & dataset-access enquiries: research@tuteliq.ai
Web: https://tuteliq.ai
Sweden
To Requestor:
[Contact information provided in Section 8]
15. Entire Agreement
This agreement constitutes the entire agreement between the parties regarding Dataset access. It supersedes all prior negotiations and understandings. No oral agreements or side arrangements are valid unless signed by authorized representatives of both parties.
16. Severability
If any provision is found invalid or unenforceable, the remaining provisions remain in full force and effect. The invalid provision shall be reformed to the minimum extent necessary to make it enforceable.
17. Signatures
For Requestor:
Organization Name: _________________________________
Primary Contact Name: _________________________________
Title: _________________________________
Email: _________________________________
Phone: _________________________________
Authorized Users List (Attach as Appendix A)
I/we certify that I/we have authority to bind this organization to this agreement, and I/we agree to all terms and conditions above.
Signature: _________________________________
Date: _________________________________
For Tuteliq AB:
Representative Name: _________________________________
Title: _________________________________
Signature: _________________________________
Date: _________________________________
Appendix A: Authorized Users
| Name | Title | Role in Project | Start Date | Notes | |
|---|---|---|---|---|---|
Maximum 5 authorized users per initial access. Additional users require prior written approval.
Appendix B: Data Handling Attestation
Requestor confirms:
- ☐ Dataset will be stored on encrypted, firewalled systems (AES-256 encryption at rest)
- ☐ Access will be restricted to named Authorized Users only
- ☐ No cloud synchronization or unencrypted storage
- ☐ Audit logs will be maintained and provided upon request
- ☐ No redistribution, training, or Prohibited Use will occur
- ☐ Breach notification procedures understood (24-hour reporting to legal@tuteliq.ai)
- ☐ Annual renewal will be completed before expiration
- ☐ Full deletion will occur upon termination
Authorized Users acknowledge receipt of this agreement:
| Name | Signature | Date | |
|---|---|---|---|
Document Version: 1.0
Effective Date: August 5, 2026
Classification: Confidential — Tuteliq Internal Use
© 2026 Tuteliq AB. All rights reserved.
- Downloads last month
- 6