wdenejko/aviai-e4b

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

wdenejko/aviai-e4b is a 7.9 billion parameter model, fine-tuned from Google's Gemma 4 E4B, specifically designed for structured decoding of aviation text. It excels at converting METAR and TAF reports into canonical JSON, and NOTAMs into category-specific extraction rows or operational classes. This model is optimized for aviation data processing, offering significantly improved exact match and value recall compared to its base model for these specialized tasks.

Loading preview...

Overview

aviai-e4b is a specialized fine-tune of Google's Gemma 4 E4B (instruction-tuned) model, developed by wdenejko. This 7.9 billion parameter model is engineered for the structured decoding of aviation text, including METAR and TAF reports into canonical JSON, and NOTAMs into category-specific extraction rows or one of 13 operational classes. It was trained as a rank-16 LoRA and merged into the base weights, providing a plain Transformers checkpoint.

Key Capabilities

  • Aviation Text Processing: Converts METAR and TAF reports into JSON, and NOTAMs into structured data or operational classes.
  • High Accuracy: Achieves significant improvements over the base Gemma 4 E4B model, with exact match scores of 94.4% for METAR to JSON, 93.3% for TAF to JSON, 82.3% for NOTAM extraction, and 95.3% for NOTAM classification.
  • Reduced Hallucination: Demonstrates lower hallucination rates compared to the base model across all aviation tasks.
  • Flexible Deployment: Available as a Transformers checkpoint (bf16) and in GGUF formats (Q8_0, f16) for llama.cpp.
  • Prompt Sensitivity: Requires exact prompt templates (provided in prompts/) for optimal performance.

Good For

  • Automated Aviation Data Processing: Ideal for applications requiring the structured extraction and interpretation of aviation weather and operational information.
  • Research and Development: A research artifact from the avtext study, suitable for further investigation into aviation text understanding.
  • Structured Data Generation: Excels at transforming unstructured aviation reports into machine-readable JSON or categorized data.

Limitations

  • Research Artifact: Not a certified aeronautical product; independent verification is required for operational or flight-safety decisions.
  • English/ICAO-format only: Trained exclusively on English and ICAO-format inputs.
  • Prompt-Sensitive: Performance is highly dependent on using the exact prompt templates it was trained on.

Finetunes
1 models