BlazingCustoms/plateclerk-decode-7b

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 28, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

BlazingCustoms/plateclerk-decode-7b is a 7.6 billion parameter LoRA fine-tune of Qwen/Qwen2.5-7B-Instruct, specifically designed to predict the offensive intent a suspicious California DMV reviewer would find in obfuscated vanity plate strings. This model excels at interpreting hidden meanings in text, reproducing human reviewer biases and inconsistencies, and is intended for entertainment and exploration rather than content moderation or decision-making about real people.

Loading preview...

Model Overview

BlazingCustoms/plateclerk-decode-7b is a 7.6 billion parameter model, fine-tuned from Qwen/Qwen2.5-7B-Instruct using LoRA (r=16, α=32). Its unique purpose is to decode offensive intent from vanity license plate strings, mimicking the interpretations of California DMV reviewers. It achieves a token-F1 score of 48.24 on a held-out test set of 1,159 rows, with all evaluations performed by a program oracle rather than an LLM judge.

Key Capabilities

  • Offensive Intent Decoding: Specialized in identifying hidden offensive meanings in obfuscated text, particularly vanity license plates.
  • Bias Reproduction: Trained to reproduce the biases and inconsistencies observed in human DMV reviewer decisions.
  • Quantization Robustness: Thoroughly evaluated across various GGUF quantizations (Q8_0, Q5_K_M, Q4_K_M), showing negligible performance degradation compared to the f16 control, with Q4_K_M even scoring slightly higher due to noise.
  • Transparent Evaluation: All evaluation metrics, graders, and test sets are published, allowing for full reproducibility and verification of reported scores.

Important Considerations

  • Emits Offensive Content: By design, the model's output includes slurs, sexual language, and other offensive strings as it directly reports the offensive interpretations it finds. It is not suitable for direct deployment to unprepared readers.
  • Not for Moderation/Decision-Making: Explicitly stated as not a content moderation system and must not be used to make decisions about real people or applications.
  • Capability Ceiling: Extensive testing with more data, better targets, RL, and increased parameters (14B) showed no significant improvement beyond the 48.24 token-F1 score, indicating a capability ceiling for this model family on this specific task.
  • Serving Path Sensitivity: Performance is highly sensitive to the serving path and prompt formatting; all reported metrics are based on /completion calls with a locally-built prompt string matching the training template.

Finetunes
1 models