jylin001206/SII-Decision-27B-LoRA

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 1, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

jylin001206/SII-Decision-27B-LoRA is a 27 billion parameter decision model, adapted from Qwen/Qwen3.8-27B using LoRA, with a context length of 32768 tokens. It is specifically optimized to return a probability distribution over user-supplied options for Choice, binary/probabilistic Noul, and ordered Score tasks, with thinking disabled. This model excels in decision-making scenarios by optimizing decision distributions rather than providing general chat supervision, achieving an 80.1517/100 score on a fixed 58-benchmark decision suite.

Loading preview...

Overview

jylin001206/SII-Decision-27B-LoRA is a 27 billion parameter decision model, part of the SII-Decision-27B family, adapted from the Qwen/Qwen3.8-27B base model. It is specifically designed to output a probability distribution across user-provided options, with its 'thinking' mechanism disabled. This model is fine-tuned for tasks requiring probabilistic decision-making rather than general conversational abilities.

Key Capabilities

  • Probabilistic Decision Output: Returns a probability distribution over given options.
  • Specialized Task Handling: Optimized for Choice, binary/probabilistic Noul, and ordered Score tasks.
  • Efficient Adaptation: Utilizes LoRA (rank 64 / alpha 128) for efficient fine-tuning, with merged weights exported for direct use.
  • Performance: Achieved a score of 80.1517 / 100 on a fixed 58-benchmark decision suite.
  • High Context Length: Supports a context length of 32768 tokens.

Good For

  • Decision-Making Applications: Ideal for systems requiring a probabilistic output for predefined choices.
  • Structured Prediction Tasks: Suitable for scenarios like binary classification (Noul) or ordered scoring where a distribution over options is needed.
  • Integration: Can be loaded directly with Hugging Face Transformers without requiring PEFT or a separate base model, supporting BF16 precision.
  • Research: The training code and data are publicly available for reproduction and further research on decision model optimization.

Finetunes
1 models