update arc2 README: H/L cycles 0/0, drop train_state note
Browse files- arc2/README.md +1 -4
arc2/README.md
CHANGED
|
@@ -22,7 +22,7 @@ resumed run).
|
|
| 22 |
|---|---|
|
| 23 |
| model | FPTRM single-z (`fp_trm_singlez`) |
|
| 24 |
| hidden size / heads / expansion | 512 / 8 / 4 |
|
| 25 |
-
| H-cycles / L-cycles |
|
| 26 |
| H-layers / L-layers | 0 / 2 |
|
| 27 |
| n_backwards_L | 6 |
|
| 28 |
| positional enc / puzzle-emb len | RoPE / 16 |
|
|
@@ -59,9 +59,6 @@ resumed run).
|
|
| 59 |
- `fp_trm_singlez.py`, `losses.py` — model and loss source for this checkpoint.
|
| 60 |
- `evaluator_ARC_step_0/submission.json` — ARC evaluation submission (predictions) for this run.
|
| 61 |
|
| 62 |
-
> Note: this folder ships **eval checkpoints only** — the `step_<N>_train_state.pt` training
|
| 63 |
-
> states (optimizer + live weights, for resuming) are not included.
|
| 64 |
-
|
| 65 |
## Reproduce the 6.2%
|
| 66 |
|
| 67 |
1. **Train** with `all_config.yaml` (FPTRM single-z, conv1d/k4, pre-norm / placement none,
|
|
|
|
| 22 |
|---|---|
|
| 23 |
| model | FPTRM single-z (`fp_trm_singlez`) |
|
| 24 |
| hidden size / heads / expansion | 512 / 8 / 4 |
|
| 25 |
+
| H-cycles / L-cycles | 0 / 0 |
|
| 26 |
| H-layers / L-layers | 0 / 2 |
|
| 27 |
| n_backwards_L | 6 |
|
| 28 |
| positional enc / puzzle-emb len | RoPE / 16 |
|
|
|
|
| 59 |
- `fp_trm_singlez.py`, `losses.py` — model and loss source for this checkpoint.
|
| 60 |
- `evaluator_ARC_step_0/submission.json` — ARC evaluation submission (predictions) for this run.
|
| 61 |
|
|
|
|
|
|
|
|
|
|
| 62 |
## Reproduce the 6.2%
|
| 63 |
|
| 64 |
1. **Train** with `all_config.yaml` (FPTRM single-z, conv1d/k4, pre-norm / placement none,
|