---
license: apache-2.0
pretty_name: Painted Wolf Decide training release open1-b7g-e4
tags:
- synthetic
- coding-agents
- tool-selection
configs:
- config_name: turn_load
  data_files:
  - split: train
    path: train.jsonl
  - split: validation
    path: val.jsonl
  - split: test
    path: holdout.jsonl
---

# Painted Wolf Decide training release open1-b7g-e4

The turn-load rows of open1-b5-e4 with their guide-unit labels relabelled from tool calls,
the companion dataset for the guide-load head. Rows, order, sanitisation, and every other
label are those of the parent release; open1-b5-e4 remains the source for the unit-rank and
code-rank inputs and the conversation exports.

| Configuration | Files | Meaning |
| --- | --- | --- |
| Turn-load | train.jsonl (2746), val.jsonl (489), holdout.jsonl (739), `corpora/independent.json` | Parent rows; `labels.guides` is whether the turn called a tool the unit attaches to or is needed with. |
| Corpus | `corpus.json` | The host's unit, tool, and skill catalog, with each unit's `attaches` and `needed_with`. |
| Trainer | `training/turn-load/` | The trainer snapshot, recipes, relabelling, probe, and audit scripts. |

`FILES.json` records the parent and released row hashes and the relabelling counts;
`SHA256SUMS` covers the release. See [Training](TRAINING.md) and `PROVENANCE.json`.