Modelpublic

sugarknight/erabi-practical-v1-experimental

sourceHugging Faceapache-2.0updated 18h agoView on Hugging Face

0likes3downloads

Model Card

ERABI Practical V1 (experimental)

This is an experimental, uncalibrated choice-ranking model. It is not an official Jev model, a validated general-purpose reasoner, or an automatic decision-maker. The model ranks 2–16 user-supplied candidate texts for a natural-language context and question and returns all candidate probabilities through the ERABI code. Decisions should be reviewed by a person.

Provenance

—Base: knowledgator/gliclass-instruct-large-v1.0, Apache-2.0, 438,672,897 parameters.
—Additional fine-tuning: one epoch on 2,414 Practical V1 training records, peak learning rate 2.5e-6, 151 optimizer steps, microbatch 2, gradient accumulation 8, fp16 AMP.
—Data: original synthetic Japanese, English, and Simplified Chinese examples in six task families: everyday arithmetic, tool choice, dialogue action, JSON conversation-log routing, reading inference, and original exam-style questions. Generated and answer-blind rejudged with DeepSeek V4.1 Flash via OrcaRouter. No actual entrance-exam questions or private chat logs were used.
—Data and training code: GitHub repository and training script. Labels remain unreviewed synthetic teacher agreement, not human gold.

Exploratory evaluation

Set	Frozen RC3 before this fine-tune	This checkpoint
Practical V1 dev, 399 cases	59.90%	77.19%
Practical V1 held-out synthetic eval, 386 cases	61.66%	76.17%
Existing RC3 Bridge, 480 cases	88.75%	88.54%

The Practical V1 eval set was used once after selecting by dev and existing-bridge results. Reading inference regressed from 54/71 to 50/71 despite aggregate gains. Candidate-order consistency on the existing bridge was 97.50%. These figures are not a benchmark of real-world correctness or Jev parity, because Practical V1 questions and labels come from the same teacher family. There is no independent human-verified final test, temperature calibration, or formal release approval for this checkpoint.

Use

bash

python -m pip install erabi
erabi predict --request request.json

The repository contains three inference formats from the same checkpoint: model.safetensors (PyTorch), onnx/fp32/model.onnx (CPU), and onnx/fp16/model.onnx (NVIDIA GPU). ERABI 0.1.2 or newer is required for --model-format; 0.1.1 uses PyTorch safetensors. Upgrade with python -m pip install --upgrade erabi. --model-format auto downloads only the selected variant: FP32 ONNX for CPU with ONNX Runtime, FP16 ONNX for CUDA with CUDA Execution Provider, and otherwise PyTorch safetensors. Install the compatible onnxruntime (CPU) or onnxruntime-gpu (GPU) separately; do not install both in one environment. You can also select --model-format pytorch, onnx-fp32, or onnx-fp16 explicitly.

The ONNX variants preserved the top-ranked choice on 90/90 unreviewed synthetic comparison cases; this is not independent human-gold validation. Experimental INT8 variants changed predictions substantially and are not distributed. The first invocation downloads the selected model; later invocations use the Hugging Face cache. Input and output JSON contracts and runtime recommendations are documented in the ERABI README. The input limit is 512 tokens; overlong inputs are rejected rather than silently truncated. The candidate probabilities are not calibrated confidence guarantees.

License and limitations

These fine-tuned weights derive from the Apache-2.0-licensed GLiClass base model and are distributed under Apache-2.0; see the Apache License 2.0 and the base model card. ERABI source code is separately MIT-licensed. Do not rely on this experimental model for high-stakes or unattended decisions.