sweetpapa/sentry-270m-moderation-v3

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.3BQuant:BF16Context Size:32kPublished:Jul 8, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

sweetpapa/sentry-270m-moderation-v3 is a 270 million parameter content-moderation classifier, distilled from a Gemma-4 31B teacher onto the FunctionGemma-270M student architecture. This model is designed to function as a fast, CPU-friendly first-pass gate in a moderation cascade, emitting a single 'moderate' function call with category, severity, and action. It excels at quickly identifying and categorizing harmful content, escalating uncertain cases to larger models for further review.

Loading preview...

Overview

sentry-270m-moderation-v3 is a compact 270M-parameter content moderation model, distilled from a Gemma-4 31B teacher. It's engineered to act as a first-pass gate in a moderation system, providing a quick classification of content into categories like hate_harassment, dangerous_content, sexual_content, violence_gore, self_harm, spam_scam, and pii_exposure.

Key Capabilities

  • Efficient Content Classification: Outputs a single moderate(...) function call with a category, severity (0.0-1.0), and policy action (allow, flag, block, escalate).
  • Low Latency: Achieves p50 latency of approximately 470 ms per call on CPU (fp32), making it suitable for real-time moderation.
  • Distilled Performance: Distilled from a larger Gemma-4 31B teacher model, balancing performance with a small footprint.
  • Specific Output Schema: Provides a structured JSON output for easy integration into automated moderation workflows.

Important Usage Notes

  • fp16 Sensitivity: Must be served in fp32 or bf16 via a transformers-native path due to numerical sensitivity. Avoid Ollama/llama.cpp/GGUF as they can silently under-flag harmful content.
  • Greedy Decoding: Recommended to use greedy decoding (do_sample=False) for optimal format validity. Avoid repetition_penalty.
  • Intended Use: Best used as a stateless first-pass filter, paired with a confidence threshold to escalate low-confidence cases to more powerful models or human review.

Limitations

  • Directional Evaluation: Evaluated against teacher labels, not a human-verified gold set.
  • Lower Recall for Hate/Harassment: hate_harassment recall is lower (~0.54) compared to other harmful categories.
  • Not for Standalone High-Stakes Decisions: Should not be used for critical decisions without secondary review or human oversight.

Finetunes
1 models