Text Classification
Transformers
Safetensors
deberta-v2
agents
agent-safety
tool-use
prompt-injection
deberta-v3
text-embeddings-inference
Instructions to use kontext-security/Merlin with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use kontext-security/Merlin with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-classification", model="kontext-security/Merlin")# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForSequenceClassification tokenizer = AutoTokenizer.from_pretrained("kontext-security/Merlin") model = AutoModelForSequenceClassification.from_pretrained("kontext-security/Merlin", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Download v3-ungated/evaluation.json from kontext-security/Merlin: direct link, hf CLI and curl.
- Browser
- Download file 1.66 kB
-
https://huggingface.co/kontext-security/Merlin/resolve/main/v3-ungated/evaluation.json
- Command line
-
hf download hf://kontext-security/Merlin/v3-ungated/evaluation.json
-
curl -L -o evaluation.json https://huggingface.co/kontext-security/Merlin/resolve/main/v3-ungated/evaluation.json
1.66 kB
| { | |
| "runtime": "merlin-v3-ungated-20261104", | |
| "weights_candidate": "merlin-v3-precision-20261007", | |
| "threshold": 0.5, | |
| "action_token_limit": 128, | |
| "action_agreement": true, | |
| "removed_gates": [ | |
| "operation_allowlist", | |
| "lexical_request", | |
| "learned_request_intent" | |
| ], | |
| "AgentDojo": { | |
| "total": 1220, | |
| "assessed": 1159, | |
| "tp": 85, | |
| "fp": 20, | |
| "positives": 352, | |
| "negatives": 868 | |
| }, | |
| "private_historical_regression": { | |
| "total": 415, | |
| "assessed": 415, | |
| "skipped": 0, | |
| "alerts": 1, | |
| "source_pool_total": 610, | |
| "excluded_before_replay": 195, | |
| "exclusion_reason": "original context could not be reliably reconstructed" | |
| }, | |
| "private_original_flagged_no_risk": { | |
| "saved": 539, | |
| "reconstructed_and_assessed": 409, | |
| "alerts": 1, | |
| "claude_code": { | |
| "saved": 133, | |
| "assessed": 128, | |
| "alerts": 0 | |
| }, | |
| "codex": { | |
| "saved": 406, | |
| "assessed": 281, | |
| "alerts": 1 | |
| } | |
| }, | |
| "private_recent_sample": { | |
| "total": 1207, | |
| "assessed": 600, | |
| "skipped": 607, | |
| "alerts": 4 | |
| }, | |
| "private_recent_alert_review": { | |
| "clear_false_alerts": 3, | |
| "uncertain_low_impact": 1, | |
| "confirmed_true_positive": 0 | |
| }, | |
| "limitations": [ | |
| "Historical examples overlap model development; regression evidence, not unseen holdout.", | |
| "AgentDojo was repeatedly consulted during development.", | |
| "Recent sample is stratified and reconstructed from pre-call transcripts; not representative of every deployment.", | |
| "Unassessed examples are not safe predictions.", | |
| "No fresh ungated ASB or AgentHarm result is claimed." | |
| ], | |
| "private_examples_published": false | |
| } | |