frontier-infra/jebadiah-9b-v2-1

Jebadiah 9B v2.1: A Specialized Decision Model

Jebadiah 9B v2.1, developed by frontier-infra, is a 9 billion parameter model specifically engineered as a decision model, not a general-purpose chat model. Built upon the Qwen/Qwen3.5-9B base, it processes typed questions and provides probabilistic outputs over predefined option labels (choice, noul, score) in a single forward pass. This model reads label-token logits directly, bypassing conversational generation.

Key Capabilities

• Probabilistic Decision-Making: Outputs probabilities for given options, suitable for structured decision tasks.
• Efficient Inference: Designed for fast, single-pass evaluation of questions with a context length of 32768 tokens.
• Specialized Training: Fine-tuned with LoRA on a diverse dataset of 16,622 records and 21,190 question presentations, focusing on choice, noul, and score question types.
• Performance Metrics: Achieved a Decision Index 0.3 public-only index of 47.20, showing gains in Knowledge & Reasoning, Language Understanding, and Retrieval & Classification compared to its predecessor.
• Optimized Inference Recipes: Can achieve higher scores (e.g., 51.53 with the T1e recipe) by employing advanced inference techniques, though at increased latency.

Good for

• Applications requiring structured decision support and probabilistic outcomes.
• Use cases where direct, non-conversational answers to typed questions are needed.
• Integrating into systems that benefit from fast, single-pass evaluation of options.
• Developers looking for a model specialized in classification, scoring, and binary (yes/no) probability tasks.