Featherless AI's Simple Jev Turns Open Hugging Face Models Into Classifier APIs
Featherless AI's open-source Simple Jev turns any Hugging Face model into a typed classifier by reading next-token logits, skipping JSON generation entirely.
- Featherless AI released Simple Jev, an Apache-2.0 open-source alternative to TypeSafe AI's closed Jev classifier.
- The server reads next-token logits for allowed answer labels instead of generating JSON, eliminating parse failures.
- Supports three question types: choice, score (fractional rubric), and noul (truth/support from 0.01 to 0.99).
- Shared-prefix KV cache reuse cuts a 4-question, 1,000-token-context request from 4,200 to 1,200 tokens of work.
- Public demo API requires no login, capped at 2k context and 2 RPS.
- Includes RFDT training scripts to distill task-specific student models with LoRA and teacher labeling.
Simple Jev turns Hugging Face models into classifier APIs
Featherless AI has released Simple Jev repository, an Apache License 2.0 library that wraps compatible Hugging Face language models in a structured classifier API. The project provides an open implementation of the interface offered by TypeSafe AI’s closed-weight Jev service.
TypeSafe Jev accepts context and returns typed decisions with confidence scores. TypeSafe lists a price of $0.042 per million input tokens, free output tokens, and response times from 70 to 500 milliseconds. Simple Jev exposes a similar contract through models such as Qwen and Gemma, with options for self-hosting, a free public demo, and a hosted Featherless endpoint.
Structured output without generation
Simple Jev derives answers from the model’s next-token logits, which represent its scores for possible next tokens. A request contains shared context and one or more questions with allowed labels. The server compares the logits for those labels and constructs the JSON response from their probabilities.
During inference, the model performs a prefill pass over the prompt and scores a small set of allowed answer tokens, such as red and blue. The request requires no autoregressive decode loop, schema-constrained sampling, or parsing of generated JSON. For choice questions, the response includes the selected label, its confidence score, and the probability distribution across candidates.
The API supports three question types:
| Type | Purpose | Result |
|---|---|---|
choice |
Selects among 2 to 50 candidate IDs. | Returns the highest-scoring candidate and the full distribution. |
score |
Rates input against an ordered rubric with 2 to 50 levels. | Returns the expected zero-based index. The value can be fractional, so a three-level rubric produces a score from 0 to 2. |
noul |
Measures truth or evidential support. | Returns a value from 0.01 to 0.99, derived from probabilities assigned to nine rating tokens. |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
