tdesk / README.md
arshy's picture
v7 roomy variant: tasks_roomy.jsonl and card section
e5bde54 verified
|
Raw History Blame Contribute Delete
2.07 kB
---
license: bsd-3-clause
pretty_name: T-Desk (OpenEnv)
tags:
- openenv
- reinforcement-learning
- agents
- terminal
configs:
- config_name: tasks
data_files:
- split: train
path: data/tasks.jsonl
- config_name: tasks_roomy
data_files:
- split: train
path: data/tasks_roomy.jsonl
---
# T-Desk
These are the Workbench task families (real-library bug fixing, reflow and downtime audits, an expense audit, and route planning), served through an interface copied field for field from OpenEnv's **`tbench2_env`**. That is the interface the OpenEnv Arena operators' smoke tests use.
**Interface**
- **Actions:** `exec`, `write_file`, `evaluate`.
- **Observations:** the instruction is repeated every turn, and output is never trimmed.
- **Reward:** binary, through `evaluate`.
- **Budget:** 4,096 completion and 8,192 context tokens.
**Pre-registered experiment.** Does training under the eval's likely interface and budget move the private score, when other environments did not? The hypothesis, predictions and verdict rules are at https://github.com/moarshy/openenv-arena-notes/blob/main/versions/v6-tdesk.md
**Calibration.** Qwen3.8-27B, thinking off, strict JSON, arena budget accounting: 13/128 ≈ 10% solved. Failures were budget exhaustion (over-reading, re-verifying) and JSON format breaks.
- **Image:** `ghcr.io/moarshy/workbench:tdesk-v1`
- **Source:** https://github.com/moarshy/workbench (`server/tdesk_environment.py`)
- **Licences:** the vendored libraries keep their own: toolz and boltons are BSD-3-Clause, tabulate and inflection are MIT.
## v7: roomy variant (`tdesk-roomy-v1`)
The same tbench2 interface with Workbench v3's budget (12,288 / 16,384) and tasks (minus `plant`). It isolates the interface from the budget after v6 scored 0/40. Pre-registered at https://github.com/moarshy/openenv-arena-notes/blob/main/versions/v7-tdesk-roomy.md
- Image: `ghcr.io/moarshy/workbench:tdesk-roomy-v1`
- Tasks: `data/tasks_roomy.jsonl`
- Calibration (strict JSON): 11/64. With room to work, format breaks, not the budget, dominate failures.