turion-ai

turion-jev-31b

README

License: apache-2.0

How it works

Prompt. Every question of a request is listed before the state. Each option gets a one-token code: A–Z, then two-letter codes that are single Gemma tokens. Up to 255 options per question are read in one pass.
Readout. The next-token log-probabilities of the option codes after the prompt, normalised over the options shown, at a temperature per question type (calibration.json: choice 1.18, yes/no 1.18, score 1.12).
Order averaging. Choice questions with at most order_avg_max options are also scored with the options reversed, and the two are averaged.
Read twice. Inputs up to 12,000 characters are repeated once in the prompt (prompt_config.json).
Attention is Gemma 4's own (causal).

Run it for the Decision Index

bash

hf download turion-ai/turion-jev-31b --local-dir /models/turion-jev-31bpip install "/models/turion-jev-31b/code[vllm]"git clone https://github.com/apolinario/decision-index && pip install -e "./decision-index[transformers]"cd decision-indexpython -m decision_index pipeline --engine turion_jev.kit_engine:TurionJev \  --option model_dir=/models/turion-jev-31b --option backend=vllm --option quantization=fp8 \  --option order_avg_max=2 --option second_pass_max_questions=1 --option repeat_max_questions=1 \  --option warm_prefix=true --option recheck=false \  --compact --out runs/turion-jev-31b

These are the options of the official run. They keep latency within the board's rule on one RTX PRO 6000:

second_pass_max_questions=1: order averaging (and recheck) only for requests with one question;
repeat_max_questions=1: read twice only for requests with one question;
warm_prefix=true: a request's shared prefix is scored once before its questions;
recheck=false: no second pass for unsure long choice questions.

The engine runs vLLM in-process, one request at a time: BF16 weights quantised online to FP8, max_model_len 65,600, prefix caching, gpu_memory_utilization 0.92. The official run used vLLM 0.32.0, torch 2.13.0+cu130 and transformers 5.19.0 (the vllm/vllm-openai image of October 2026) with the kit at 25c6ec7.

Declared limit: prompts up to 65,536 tokens. A longer request is refused (Unsupported), never truncated. In the public 0.3 suite that is 226 ToolRet requests and 1 BRIGHT request.

A TypeSafe-compatible POST /v1/systemone server is also included (python -m turion_jev.serve --model /models/turion-jev-31b --backend vllm --quantization fp8). It takes fewer options than the kit engine; use the engine for the settings above.

Training

Method. LoRA r64 (alpha 128) on all linear layers, learning rate 2e-5, 450 steps of ~65k tokens, sequences up to 16,384 tokens, options shuffled, inputs up to 12,000 characters repeated once. The loss is cross-entropy against soft (probability) targets over the option codes.
Merge. A uniform soup of the checkpoints at steps 300, 375 and 450, merged into the base at full scale and saved in BF16.
Calibration. One temperature per question type, fitted on held-out data.
Compute. A training job on the Turion AI model training platform: one NVIDIA B300, about 7.3 hours.

Data

157,106 training rows:

tasksource/tasksource-jev-typed-decisions: 106,680 rows from 439 source datasets, keeping only sources tagged for commercial use, plus iSarcasmEval (MIT) and ContractNLI (CC BY 4.0), whose original releases allow it.
About 24.5k synthetic decision tasks. They were written by Gemma-4-31B-it and labelled by Gemma-4-31B-it in thinking mode. The most uncertain 12% were labelled again by DeepSeek V4.1 Flash, whose API terms allow distillation.
SargeDev/jev-distill-corpus-v3 (Apache-2.0).
Annotator distributions as soft labels: google/civil_comments (CC0), ucberkeley-dlab/measuring-hate-speech (CC BY 4.0) and sentence-transformers/stsb (CC BY-SA 4.0).

The Decision Index suite was not trained on. Every training row was checked against the 0.3 suite by line match, 13-gram overlap and exact match, and 4,395 rows were dropped. Evaluation sets next to the suite's benchmarks (VAST, ANLI, MedNLI and others) were excluded.

Limitations

Prompts over 65,536 tokens are refused.
Attention is causal; the model does not read the prompt bidirectionally.
The reported run used FP8. The weights are BF16.
Probabilities are calibrated on our held-out data. Check calibration on your own data before relying on the numbers.

Licence

Apache-2.0, like the base model; see LICENSE and NOTICE. The training data sources keep their own licences, listed above.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

turion-ai

Model Tree

Base

google/gemma-4-31B-it

Fine-tuned

this model

Input Modalities

TextImage

Output Modalities

Text

Supported Functionality