Datasets:
context-retention-control-v1
9,834 original typed rows; 400 complete synthetic families. Original generated content and labels are CC0-1.0. Source code keeps the repository's MIT license.
Prepared and independently audited only. Not used for training or model evaluation. This addition changes neither the frozen release-v2 mixture nor the data used by the published 2B/9B checkpoints.
Keep or discard eligible completed tool-call records and full outputs under an explicit fixed retention policy. The shared state contains context, goal and history; full tool outputs are omitted. Two Noul questions are built per eligible call. Labels follow visible goal dependency closure, exact-evidence needs and output recoverability. Pinned and pending calls are software gates, without model labels. Whole task graphs and their goal/recoverability counterfactuals share a split. OOD reserves diamond dependencies and wording. This finite synthetic task is not evidence of useful arbitrary-session compaction.
| Split | Typed rows |
|---|---|
| train | 6,138 |
| calibration | 456 |
| validation | 558 |
| test | 522 |
| ood | 2,160 |
The community interface reference supplies the task shape. The task instances, labels and instructions are independently authored. No original third-party transcript, whole documentation page or model response is bundled.
Use only state, question, kind and options as model inputs. Targets are reference labels; metadata and auxiliary cases may contain privileged information. The Parquet columns and JSON decoding rules match the root dataset card. Decompressing each raw split recovers its original JSONL bytes exactly. Whole-family splits include correlated counterfactuals; row counts must not be described as independently collected real-world cases.
From the source snapshot's root, regenerate in a new directory with:
python3 -m jev.context_retention_data --output-dir data/context-retention-control-v1 --groups 400 --ood-groups 80 --seed 942
No trained-model accuracy, useful compaction, natural-video classification, or general API failure-detection performance is claimed by this release.