| # INTACT Unified E5 |
|
|
| This directory contains the complete controlled **four-task shared-encoder E5** |
| checkpoint matrix from the INTACT paper. A shared encoder is trained jointly on |
| PushT, Cube, Reacher, and TwoRoom, with task-specific lightweight heads. |
|
|
| ## Controlled comparison |
|
|
| | Cell | Archive prefix | Native interface | Official macro SR | |
| |---|---|---|---:| |
| | LeWM | `paper-e5-lewm-*` | CEM 300x30 | 66.17 | |
| | Inverse only | `paper-e5-inverse-only-*` | Direct | 68.19 | |
| | Goal intent only | `paper-e5-goal-intent-only-*` | Direct | 84.92 | |
| | **Goal-displacement INTACT** | `intact-goal-e5-*` | **Direct** | **91.22** | |
|
|
| The Goal-displacement INTACT task-level result is: |
|
|
| | PushT | Cube | Reacher | TwoRoom | Macro SR | |
| |---:|---:|---:|---:|---:| |
| | 86.11 +/- 0.96 | 100.00 +/- 0.00 | 97.22 +/- 0.84 | 81.56 +/- 3.67 | **91.22 +/- 0.51** | |
|
|
| Success rates are percentages averaged over training seeds `0/42/3072`. |
| Each checkpoint is evaluated on 100 episodes for evaluation seeds `0/1/42`; |
| `+/-` is the sample standard deviation across training-seed means. |
|
|
| These values use the corrected causal-continuation evaluator: the initial |
| actor history contains only strictly pre-start actions, and each action |
| executed by the evaluated policy is appended exactly once. The current values |
| are recorded in `AUDITED_RESULTS_20260914.json`. |
|
|
| ## Contents |
|
|
| - Six archived training cells and three training seeds. The current comparison |
| above reports the four goal-displacement cells used by the revised analysis. |
| - Four task-specific checkpoint shards in each archive. |
| - `6 x 3 x 4 = 72` checkpoints in total. |
| - `AUDITED_RESULTS_20260914.json`: authoritative corrected-evaluator SR values. |
| - `PAPER_E5_MATRIX_MANIFEST.json`: checkpoint inventory and extracted shard hashes. |
| - `PAPER_E5_GOAL_MANIFEST.json`: Goal-displacement INTACT checkpoint metadata. |
| - `SHA256SUMS`: archive integrity records. |
| - `RUNTIME_SHA256SUMS`: frozen evaluator runtime fingerprint. |
|
|
| ## Download |
|
|
| ```bash |
| hf download INTACT-JEPA/INTACT --include "INTACT-unified/*" \ |
| --local-dir checkpoints-release |
| |
| cd checkpoints-release/INTACT-unified |
| sha256sum -c SHA256SUMS |
| ``` |
|
|
| These paper checkpoints use the frozen paper runtime and legacy Actor checkpoint |
| grammar. Use the paper evaluation wrappers supplied by the INTACT source release; |
| the weights are intentionally separate from the current clean E1 runtime. |
|
|
| The immutable original release remains available at |
| [`paper-e5-goal-v1`](https://huggingface.co/INTACT-JEPA/INTACT/tree/paper-e5-goal-v1). |
|
|
