Hagrun/moddog-l1-safety-qwen2.5-3b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 16, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Hagrun/moddog-l1-safety-qwen2.5-3b is a 3.1 billion parameter content moderation classifier, fine-tuned from Qwen2.5-3B-Instruct, designed for fast, first-pass triage of Discord-style chat messages. It returns structured JSON verdicts (flag/category/confidence/reason) and is specifically calibrated to be honestly uncertain on hard cases, escalating them to larger models rather than mis-judging. This model excels at identifying toxicity, harassment, hate speech, sexual content, self-harm, and violence, serving as the L1 fast-triage tier in the ModDog pipeline.

Loading preview...

ModDog L1 — Qwen2.5-3B Moderation Judge

This model is a 3.1 billion parameter content moderation classifier, fine-tuned from Qwen2.5-3B-Instruct, specifically for Discord-style chat. It functions as the L1 fast-triage tier in the ModDog pipeline, providing structured JSON verdicts (flag / category / confidence / reason). A key design principle is its intentional uncertainty on difficult cases, ensuring they are escalated for human review or larger models rather than being confidently misclassified.

Key Capabilities & Features

  • Fast First-Pass Triage: Optimized for rapid assessment of chat messages.
  • Specific Moderation Categories: Judges toxicity, harassment, hate_speech, sexual_content, self_harm, and violence (plus benign).
  • Structured JSON Output: Provides detailed verdicts for automated processing.
  • Calibration-First Training: Deliberately includes borderline examples to prevent overconfident false positives.
  • Prompt-Injection Resistance: Trained to ignore embedded instructions and judge message content.
  • Clean Lineage: Trained on CC0 and human-curated data, ensuring an Apache-2.0 compatible lineage.
  • Production-Grade Weights: This release contains the exact weights used in ModDog production since 2026-06-23.

Important Considerations

  • Not for Deterministic Rules: This model is not trained for spam, profanity, or other pattern-matchable violations; these should be handled by a deterministic rule layer.
  • Calibration Note: The model's verbalized confidence field is not inherently well-calibrated. Users should fit their own calibrator if thresholding on confidence.
  • Prompt Sensitivity: Requires a detailed system rubric with few-shot examples (around 1.9k tokens) at inference time to maintain performance and calibration. Short prompts can significantly degrade accuracy.
  • Domain Specificity: Tuned for Discord chat; may over-flag benign-but-edgy content at low confidence on out-of-distribution text.

Finetunes
1 models