Original title: perplexity-ai/pplx-decider-v1.3
Uploaded by Perplexity on 11 9, this is the next version of v1.2, a multimodal classification model based on Qwen3.8-27B. A request provides a state (text, JSON, or image) and several questions, with each question being one of four types: choice (1–255 options, returning the probability of each option), noul (yes/no, returning the probability of true), score (2–255 scale rating, returning the expected level), and the newly added locate (extracting verbatim snippets from the original text, with character positions and scores). It does not use full vocabulary output, but rather a 255-row readout head, with full attention layers having bidirectional visibility to the input; the temperature has been calibrated on held-out data, and decision_config.json is present. Apache-2.0.
Why it matters Suitable for high-volume decision tasks such as ticket routing, content moderation, and routing decisions; it only reads probabilities without generating text, facilitating ranking and threshold setting, under the Apache license
Rising score 27 · Likes 92 · Downloads 1304
Reasons
- In the default separate mode, each question is computed independently, so adding, removing, or reordering other questions does not affect the answer; the isolated mode shares a single read-in for greater efficiency, with the official documentation stating that 1.5–2.2% of preferred answers change on samples
- With fast=True, the official documentation states that separate mode is approximately 1.7–1.9 times faster, and isolated mode is 9–18 times faster, with the preferred answer matching the reference implementation at a rate of 99.4% or higher
- The official documentation acknowledges the tasksource dataset, stating that most improvements come from it; training, evaluation, and use cases are detailed in Perplexity's blog
Usage
- Requires Python 3.12+ and a CUDA GPU; BF16 weights take approximately 49 GiB plus working memory; after snapshot_download, add source/src to sys.path and use from pplx_decider import Decider; alternatively, use pip install ./source
- After Decider.from_pretrained(checkpoint, device="cuda"), call decide({"state": ..., "questions": {...}, "images": [...]}); for ranking, use decider.logits() to get the difference between true and false for noul, as probabilities do not saturate into ties
- locate requires a plain text state without images, with a maximum of 128 questions per request; images are automatically resized to within 1.6 MP, occupying approximately 1,600 positions; the default max_length is 32,768 tokens
Risks
- The card states "the repository is private and requires a token with permissions," but the repository was publicly readable at the time of this review; actual access may vary
- The card does not include evaluation scores; standard causal inference cannot reproduce this checkpoint, requiring the use of the provided code, and prompt construction for v1.1 / v1.2 cannot be reused
- When relying on probability values, the official documentation recommends recalibrating with your own traffic, as a single temperature is not suitable for all tasks. This entry has not been run