| ftb-sciworld-repro |
| ================== |
|
|
| A runnable ScienceWorld reproduction bundle. It redistributes two upstream |
| Apache-2.0 projects unmodified, and adds configuration and scripts on top. |
|
|
| Everything in this bundle is licensed under the Apache License, Version 2.0. |
| See LICENSE. |
|
|
|
|
| Redistributed upstream code |
| --------------------------- |
|
|
| 1. tcod/ |
| TCOD -- Temporal Curriculum in On-Policy Distillation |
| https://github.com/kokolerk/TCOD |
| Copyright the TCOD authors. Licensed under Apache-2.0; see tcod/LICENSE. |
|
|
| Vendored at commit 17a8af222017fb78a44a50e5711bfa08543de36b, exported with |
| `git archive`, so the tree is byte-identical to that commit. NOT MODIFIED in |
| this bundle as shipped. |
|
|
| Note that scripts/setup.sh and scripts/run.sh copy overlay/trinity/ over |
| tcod/trinity/ at install and launch time. After you run either script, the |
| working copy of tcod/ is no longer pristine: the ScienceWorld workflow files |
| under tcod/trinity/common/workflows/envs/TCOD/ are replaced by the overlay |
| below. Re-download the bundle to recover the unmodified tree. |
|
|
| 2. overlay/ |
| FutureBridge-OPD (FTB) -- "Look Ahead Before You Distill" |
| https://github.com/ChenChiShui/FutureBridge-OPD |
| Copyright the FutureBridge-OPD authors. Licensed under Apache-2.0. |
|
|
| The `source/` directory of commit bfcdf91, exported with `git archive`. |
| NOT MODIFIED. This is the same overlay that the upstream release's own |
| scripts/install_overlay.sh installs. |
|
|
| 3. configs/ftb.yaml |
| Copied verbatim from FutureBridge-OPD configs/scienceworld/ftb.yaml at the |
| same commit. NOT MODIFIED. |
|
|
|
|
| Added in this bundle |
| -------------------- |
|
|
| The following files are new work by the bundle author, not from either upstream |
| project: |
|
|
| configs/opd.yaml |
| configs/guided_opd.yaml |
| configs/tcod_b2f.yaml |
| configs/tcod_f2b.yaml |
| Baseline configurations. The upstream release ships ScienceWorld configs |
| for FutureBridge and its ablations only; for the four baselines it ships |
| the workflow implementations but no configuration, giving instead a |
| "Baseline workflow mappings" table in its configs/README.md. These four |
| files are a reconstruction from that table, holding every shared setting |
| identical to ftb.yaml. Values the table does not determine are documented |
| in the header of each file and in README.md. |
|
|
| configs/bench.yaml.tmpl |
| An evaluation configuration. The upstream release performs no evaluation; |
| see README.md and scripts/eval.sh. |
|
|
| scripts/setup.sh |
| scripts/prepare_data.py |
| scripts/run.sh |
| scripts/eval.sh |
| scripts/run_slurm.sbatch |
| README.md |
| NOTICE |
| requirements.txt (copied from FutureBridge-OPD, unmodified) |
|
|
|
|
| Data |
| ---- |
|
|
| scripts/prepare_data.py downloads the ScienceWorld task split from |
| https://huggingface.co/datasets/SeanWang0027/scienceworld-tcod-split |
| That dataset mirrors the split produced by TCOD's own |
| TCOD_examples/scienceworld/get_sciworld_data.py, with the machine-local |
| `jar_path` field blanked so the pip `scienceworld` package's bundled jar is used |
| instead. No task content is altered. |
|
|
| ScienceWorld itself (https://github.com/allenai/ScienceWorld) is installed from |
| PyPI by scripts/setup.sh and is not redistributed here. |
|
|
