---
license: mit
tags:
  - spare
  - tool-use
  - self-play
  - rl-environments
pretty_name: SPARE tool-use glory runs — generated environments
---

# SPARE tool-use "glory" runs — generated environments

All generated environments and visualizer data for the two longest SPARE tool-use
self-play training runs (399 training iterations each, June–July 2026):

| Run | Model | Steps | run_id |
|---|---|---|---|
| `0701_tooluse_glory_kl0` | Qwen3-4B-Instruct-2507 | iter 399 | `223t1pws` |
| `0705c_glory_r8_premerge_30b` | Qwen3-30B-A3B-Instruct-2507 | iter 399 | `5hvg1dna` |

SPARE trains a model in dual roles: as **environment proposer** (generating
tool-use games as Python `ToolUseBaseEnv` subclasses) and as **actor** (playing
them). Skills covered: `api_orchestration`, `data_retrieval`,
`state_modification`, `error_recovery`, `tool_selection`,
`multi-step_workflows`.

## Contents

### `raw_envs/`

One tarball per run with the raw generated environment files from training:

- `spare_games/` — accepted envs at the final step (`game_<step>_<idx>_<skill>.py`)
- `spare_games_cache/step_XXXXX/` — per-step snapshots across training, each env
  with its `_hint.txt` hint file (the 30B run has 115 snapshots up to step 392)
- `spare_games_rejected/` — every rejected generation with a paired
  `.reason.txt` validator explanation

### `site/`

Self-contained payload for the SPARE Run Visualizer
(https://spare.dobby-orchestrator.uk), filtered to these two runs:

- `data.js` — sets `window.REPORT_DATA` (run index: steps, games, metrics).
  Drag & drop onto the visualizer, or serve next to its `index.html`.
- `v2/g/<rid>/<game_id>.json` — per-game detail files the visualizer fetches on
  demand: full `game_code`, proposer prompt/response/thinking, and solver
  trajectories. 4,373 games total (4,023 from the 30B run, 350 from the 4B run).

## Provenance

Extracted from training runs at `spare_runs/0701_tooluse_glory_kl0/20260701_112938`
and `spare_runs/0705c_glory_r8_premerge_30b/20260705_074459` (MSR SPARE project).