You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

These files contain the Decision Index suite rows, including GPQA and HLE text. Do not redistribute them or use them for training.

Log in or Sign Up to review the conditions and access this dataset content.

Decision Index results: Decider chat readout

Full runs of the Decision Index suite for a readout from decider. The readout is one forward pass per request. The softmax is taken over the option-letter logits and then divided by a temperature.

run checkpoint temperature Decision Index
runs/full-chat-gemma4-31b-it-adaptiveT google/gemma-4-31B-it @ 842da379, bf16 T(n) = 10.124 − 1.633 ln n (n = number of options) 59.84
runs/full-chat-qwen3.6-27b-T1.943 Qwen/Qwen3.6-27B @ 6a9e13bd, bf16 1.943 55.73
run edition, kit checkpoint temperature balanced_skill
runs/full021-decider-12b-v2-T2.45 0.2.1, 87d4650 (150,759 requests) Mapika/decider-12b @ 8ac1efa7 (Gemma-4-12B-it + merged LoRA), bf16 2.45 50.41

The first two runs are edition 0.1 (132,422 requests, kit 52a6989) on stock checkpoints; their score is the 0.1 index. The decider-12b run is edition 0.2.1 on a fine-tuned checkpoint. Its NOTES.md states how it was served: 7 long requests were read with vLLM, and the file gives the agreement check.

Each run folder holds the kit's scoring files (scores.json, index.json, benchmark-summary.json) and results.jsonl.gz, which keeps the last record per run_id. It also holds the raw per-shard kit output with every error record under shards/, the launcher log and the server logs, and the serving scripts and temperature-fit records under code/. All three results files re-score with python -m decision_index score (with --edition 0.2.1 for the third) to the stored scores.json.

Access is gated because the rows carry the suite's GPQA and HLE text. Requests are approved manually.

Downloads last month
49