Looped Fast-MCTD trained models / 已完成模型归档
Snapshot: 2026-09-15 (UTC). This release covers completed training runs in the idea-loop-mcdt project.
| Category | Models | Training | Saved steps |
|---|---|---|---|
| 01 Fast-MCTD baselines | 7 | Full training | 200000 and 200005 |
| 02 Historical Loop Flow v1 | 6 | Full training; historical/negative or unevaluated variants | 200000 and 200005 |
| 03 Elastic Loop v2 | 5 | Full refiner training on frozen baselines | 200000 and 200005 |
| 04 Elastic v3 development pilots | 8 | Completed 5000-step warm-start pilot budget | 5000 |
26 distinct trained models, 44 full-state checkpoints. Each model ZIP is independently downloadable. Primary evaluation uses step 200000 (step 5000 for development pilots). Final step 200005 is retained for archival/resume purposes.
本次包含本项目全部已完成训练模型的主评估/最终检查点，不包含每1000步的历史中间快照。最新full11任务仍在训练/等待的模型、提前停止的旧实验和几步的preflight测试不作为已完成模型上传；清单中的excluded逐项记录。v3的5000步是完成计划的微调试验，不是从零训练到200000步。历史Loop Flow不得解读为新方法有效性的证明。作者发布的预训练模型和DQL控制器不是我们训练的模型，不计入也不重复发布。
Download and load
This is a public model archive. Anyone can download the files without logging in or requesting access.
hf download guo925658/looped-fast-mctd --local-dir ./looped-fast-mctd-models
Or select an individual ZIP in Files and versions. These are custom PyTorch models; loading uses the bundled code, rather than a Transformers AutoModel class. All 44 checkpoints passed strict CPU model loading using the bundled source and relocated dependency paths.
- Download
source-and-guides.zip,manifest.json,SHA256SUMS, and the desired model ZIPs. For Elastic models also download everydependency_archivesZIP listed in the manifest: the matching baseline, and for v3 its v2 warm-start model. - Verify downloads with
sha256sum --ignore-missing -c SHA256SUMS. - Extract all model ZIPs into one writable directory (the bundle root), preserving
checkpoints/<run>/<arm>/...andmodel_configs/.... Putmanifest.jsonin that same root. - Extract
source-and-guides.zip. In itssource/directory, set up the pinned Python 3.10 / Torch 2.5.1 environment described indocs/quickstart.md(including the bundled OGBench wheel). - CPU load example, from
source/:
python -m scripts.load_published_model --bundle-root /path/to/downloaded-models --run repro_large_v2 --arm baseline
python -m scripts.load_published_model --bundle-root /path/to/downloaded-models --run elastic_v3_cube_double_gated --arm elastic_loop
The helper verifies the selected checkpoint SHA256, relocates dependency paths, verifies their hashes through the native builder, and performs strict model loading. In Python, from scripts.load_published_model import load_model returns (model, config). Use device="cuda" on a supported GPU. Different native backends must be loaded in separate Python processes.
Checkpoints remain byte-for-byte identical to the originals and include model, optimizer, scaler, sampler, RNG, configuration and counters. They require torch.load(..., map_location="cpu", weights_only=False) because they include full training state; only load trusted files. Original configs retain historical absolute paths for provenance; the helper remaps dependencies in memory without rewriting checkpoint bytes. Existing native evaluation CLI performs exact original-config checks, so moving an archive alone does not make that CLI portable; use the provided model loader for loading, and adapt evaluation paths deliberately.
Full environment evaluation also needs official OGBench datasets and, for Ant/Cube, the released DQL controllers. See bundled docs/quickstart.md, docs/hard_tasks.md, docs/full11_protocol.md and upstream READMEs for assets and evaluation protocols. This archive validates model loading; it does not assert a fresh end-to-end evaluation on another machine. Historical metrics are in docs/EXPERIMENT_SUMMARY_ZH.md.
Provenance
manifest.json: checkpoint hashes, steps, tasks, dependencies, and exclusions.
source-manifest.json inside the source ZIP: exact bundled source hashes (including the export/load helpers) and source commit.
VALIDATION.json: local readability, finite tensors, dependency hashes, and strict loading results.
Licenses and third-party notices are preserved in the source ZIP.
Model downloads
| Family | Task | Run | Model archive | Required model archives |
|---|---|---|---|---|
| 01_fast_mctd_baselines | antmaze-giant-navigate-v0 | hard_antmaze_giant_v1 | Download ZIP | None |
| 01_fast_mctd_baselines | antmaze-large-navigate-v0 | hard_antmaze_large_v1 | Download ZIP | None |
| 01_fast_mctd_baselines | cube-double-play-v0 | hard_cube_double_v1 | Download ZIP | None |
| 01_fast_mctd_baselines | cube-triple-play-v0 | hard_cube_triple_v1 | Download ZIP | None |
| 01_fast_mctd_baselines | pointmaze-giant-navigate-v0 | repro_giant_v2 | Download ZIP | None |
| 01_fast_mctd_baselines | pointmaze-large-navigate-v0 | repro_large_v2 | Download ZIP | None |
| 01_fast_mctd_baselines | pointmaze-medium-navigate-v0 | repro_medium_v2 | Download ZIP | None |
| 02_legacy_loop_flow | antmaze-large-navigate-v0 | hard_antmaze_large_v1 | Download ZIP | None |
| 02_legacy_loop_flow | cube-double-play-v0 | hard_cube_double_v1 | Download ZIP | None |
| 02_legacy_loop_flow | cube-triple-play-v0 | hard_cube_triple_v1 | Download ZIP | None |
| 02_legacy_loop_flow | pointmaze-giant-navigate-v0 | linear_flow_giant_v1 | Download ZIP | None |
| 02_legacy_loop_flow | pointmaze-large-navigate-v0 | linear_flow_large_v1 | Download ZIP | None |
| 02_legacy_loop_flow | pointmaze-medium-navigate-v0 | linear_flow_medium_v1 | Download ZIP | None |
| 03_elastic_v2 | antmaze-giant-navigate-v0 | elastic_v2_antmaze_giant | Download ZIP | 01_fast_mctd_baselines__hard_antmaze_giant_v1__baseline.zip |
| 03_elastic_v2 | antmaze-large-navigate-v0 | elastic_v2_antmaze_large | Download ZIP | 01_fast_mctd_baselines__hard_antmaze_large_v1__baseline.zip |
| 03_elastic_v2 | cube-double-play-v0 | elastic_v2_cube_double | Download ZIP | 01_fast_mctd_baselines__hard_cube_double_v1__baseline.zip |
| 03_elastic_v2 | cube-triple-play-v0 | elastic_v2_cube_triple | Download ZIP | 01_fast_mctd_baselines__hard_cube_triple_v1__baseline.zip |
| 03_elastic_v2 | pointmaze-large-navigate-v0 | elastic_v2_pointmaze_large | Download ZIP | 01_fast_mctd_baselines__repro_large_v2__baseline.zip |
| 04_v3_pilots | antmaze-giant-navigate-v0 | elastic_v3_antmaze_giant_control | Download ZIP | 01_fast_mctd_baselines__hard_antmaze_giant_v1__baseline.zip, 03_elastic_v2__elastic_v2_antmaze_giant__elastic_loop.zip |
| 04_v3_pilots | antmaze-giant-navigate-v0 | elastic_v3_antmaze_giant_gated | Download ZIP | 01_fast_mctd_baselines__hard_antmaze_giant_v1__baseline.zip, 03_elastic_v2__elastic_v2_antmaze_giant__elastic_loop.zip |
| 04_v3_pilots | antmaze-giant-navigate-v0 | elastic_v3_antmaze_giant_gated_quality | Download ZIP | 01_fast_mctd_baselines__hard_antmaze_giant_v1__baseline.zip, 03_elastic_v2__elastic_v2_antmaze_giant__elastic_loop.zip |
| 04_v3_pilots | antmaze-giant-navigate-v0 | elastic_v3_antmaze_giant_quality | Download ZIP | 01_fast_mctd_baselines__hard_antmaze_giant_v1__baseline.zip, 03_elastic_v2__elastic_v2_antmaze_giant__elastic_loop.zip |
| 04_v3_pilots | cube-double-play-v0 | elastic_v3_cube_double_control | Download ZIP | 01_fast_mctd_baselines__hard_cube_double_v1__baseline.zip, 03_elastic_v2__elastic_v2_cube_double__elastic_loop.zip |
| 04_v3_pilots | cube-double-play-v0 | elastic_v3_cube_double_gated | Download ZIP | 01_fast_mctd_baselines__hard_cube_double_v1__baseline.zip, 03_elastic_v2__elastic_v2_cube_double__elastic_loop.zip |
| 04_v3_pilots | cube-double-play-v0 | elastic_v3_cube_double_gated_quality | Download ZIP | 01_fast_mctd_baselines__hard_cube_double_v1__baseline.zip, 03_elastic_v2__elastic_v2_cube_double__elastic_loop.zip |
| 04_v3_pilots | cube-double-play-v0 | elastic_v3_cube_double_quality | Download ZIP | 01_fast_mctd_baselines__hard_cube_double_v1__baseline.zip, 03_elastic_v2__elastic_v2_cube_double__elastic_loop.zip |