card: diagrams; credit Pranav Ramakrishnan by name
Browse files- .gitattributes +3 -0
- README.md +7 -1
- assets/how.png +3 -0
- assets/results.png +3 -0
- assets/train.png +3 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
assets/how.png filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
assets/results.png filter=lfs diff=lfs merge=lfs -text
|
| 38 |
+
assets/train.png filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
|
@@ -36,8 +36,12 @@ Code, training pipeline and paper: [github.com/a3ro-dev/reflex-2048](https://git
|
|
| 36 |
|
| 37 |
This is not a record. It's a small, cheap data point.
|
| 38 |
|
|
|
|
|
|
|
| 39 |
## Results
|
| 40 |
|
|
|
|
|
|
|
| 41 |
10,000 fresh games, greedy argmax policy, end-of-schedule checkpoint (not selected by eval score):
|
| 42 |
|
| 43 |
| mean | median | max | 2048 | 4096 | 8192 | 16384 | 32768 | 65536 |
|
|
@@ -101,6 +105,8 @@ The checkpoint is a plain `torch.save` dict (`cfg`, `model` state dict, `step`,
|
|
| 101 |
|
| 102 |
## Training
|
| 103 |
|
|
|
|
|
|
|
| 104 |
- **Loss:** cross-entropy to the teacher's move plus a 0.25× Huber loss on log return-to-go. Random D4 symmetry per sample.
|
| 105 |
- **Data:** 360M depth-3 and 100M depth-4 teacher self-play decisions.
|
| 106 |
- **Schedule:** 261M samples total. The model starts from a first run's step-6000 checkpoint, taken before any DAgger data reached the trainer. It then trains 212M samples on depth-3/depth-4 data at 50/50 with AdamW, peak lr 5e-4 and a 1.9 h cosine schedule, in bf16 on one A100.
|
|
@@ -109,4 +115,4 @@ The teacher is GPL-3.0 code. The repo ships only a patch script for it, not the
|
|
| 109 |
|
| 110 |
## Thanks
|
| 111 |
|
| 112 |
-
The RunPod credits that paid for all the compute came from [
|
|
|
|
| 36 |
|
| 37 |
This is not a record. It's a small, cheap data point.
|
| 38 |
|
| 39 |
+

|
| 40 |
+
|
| 41 |
## Results
|
| 42 |
|
| 43 |
+

|
| 44 |
+
|
| 45 |
10,000 fresh games, greedy argmax policy, end-of-schedule checkpoint (not selected by eval score):
|
| 46 |
|
| 47 |
| mean | median | max | 2048 | 4096 | 8192 | 16384 | 32768 | 65536 |
|
|
|
|
| 105 |
|
| 106 |
## Training
|
| 107 |
|
| 108 |
+

|
| 109 |
+
|
| 110 |
- **Loss:** cross-entropy to the teacher's move plus a 0.25× Huber loss on log return-to-go. Random D4 symmetry per sample.
|
| 111 |
- **Data:** 360M depth-3 and 100M depth-4 teacher self-play decisions.
|
| 112 |
- **Schedule:** 261M samples total. The model starts from a first run's step-6000 checkpoint, taken before any DAgger data reached the trainer. It then trains 212M samples on depth-3/depth-4 data at 50/50 with AdamW, peak lr 5e-4 and a 1.9 h cosine schedule, in bf16 on one A100.
|
|
|
|
| 115 |
|
| 116 |
## Thanks
|
| 117 |
|
| 118 |
+
The RunPod credits that paid for all the compute came from [Pranav Ramakrishnan](https://www.linkedin.com/in/pranavram98/). Thank you.
|
assets/how.png
ADDED
|
Git LFS Details
|
assets/results.png
ADDED
|
Git LFS Details
|
assets/train.png
ADDED
|
Git LFS Details
|