File size: 2,516 Bytes
5e635ed 7f214fa 5e635ed 7f214fa 5e635ed 7f214fa 5e635ed 7f214fa 5e635ed 7f214fa 5e635ed | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 | # 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).
|