MeridianSafety/MERIT-XS-Preview

MERIT-XS (Research Preview)

MERIT-XS is an early multilingual moderation encoder developed by Meridian Safety for research into compact, Unicode-aware safety classification systems.

This release packages a binary toxicity research preview built from:

• MERIT-XS encoder pretraining
• a top-2-layer moderation adaptation run
• a binary moderation head

This artifact is not production-ready and should not be used as a standalone safety system.

Included files

• merit_xs_preview.pt
  • exported moderation artifact with adapted encoder weights and binary head
• infer_merit_xs.py
  • CLI inference entrypoint
• load_merit_xs.py
  • simple Python loader for local use
• metrics_summary.json
  • dev/test metrics and threshold sweep summary
• profile_summary.json
  • lightweight export-run timing summary
• merit/
  • local model package
• assets/tokenizers/merit/
  • tokenizer files

Setup

pip install -r requirements.txt

License

This package uses the included LICENSE.txt:

• MERIT Research Preview License (MRPL v1.0)
• research, evaluation, and benchmarking use are allowed
• commercial deployment and hosted/public API use require separate permission

CLI usage

python infer_merit_xs.py \
  --text "you are awful" \
  --text "thanks for your help"

You can also pass an explicit checkpoint path:

python infer_merit_xs.py \
  --checkpoint merit_xs_preview.pt \
  --text "you are a stupid idiot"

Python usage

from load_merit_xs import load_merit_xs

model = load_merit_xs()
results = model.predict(
    [
        "you are awful",
        "thanks for your help",
        "you are a stupid idiot",
    ]
)
print(results)

Output schema

Each prediction returns:

• score
  • sigmoid(logit)
• decision
  • allow | review | action
• confidence
  • threshold-distance heuristic only
• decision_band
  • same band label used for the decision

Important: confidence here is a preview-time decision-margin style heuristic, not calibrated probability confidence.

Current limitations

• Binary toxicity preview only
• Not a full moderation taxonomy
• Weak coverage for some safety categories, including self-harm / threat-style language
• Message-level only
• Incomplete multilingual and adversarial evaluation

Research note

This package is intended for:

• research
• benchmarking
• representation-transfer experiments
• moderation evaluation

It is not intended for:

• production moderation
• safety-critical enforcement
• fully automated policy decisions

Existing model card

This package is being prepared for the existing Hugging Face repo:

MeridianSafety/MERIT-XS-Preview