Upload AtomSLM checkpoint
Browse files- README.md +72 -3
- config.json +12 -0
- pytorch_model.bin +3 -0
- tokenizer/tokenizer.json +0 -0
README.md
CHANGED
|
@@ -1,3 +1,72 @@
|
|
| 1 |
-
---
|
| 2 |
-
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: pytorch
|
| 3 |
+
tags:
|
| 4 |
+
- atomslm
|
| 5 |
+
- language-model
|
| 6 |
+
- shared-weights
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
# AtomSLM-600K
|
| 10 |
+
|
| 11 |
+
**AtomSLM** is a compact, shared-weight language model family built on the
|
| 12 |
+
AtomNet architecture (Shared Weight Core + Per-Layer FiLM Modulation).
|
| 13 |
+
|
| 14 |
+
## Model Details
|
| 15 |
+
|
| 16 |
+
| Field | Value |
|
| 17 |
+
|---------------|---------------------------|
|
| 18 |
+
| Architecture | AtomNet (shared-core FiLM)|
|
| 19 |
+
| Parameters | 0.613M |
|
| 20 |
+
| Vocab size | 4096 |
|
| 21 |
+
| d_model | 120 |
|
| 22 |
+
| Layers | 8 |
|
| 23 |
+
| FFN mult | 2.0 |
|
| 24 |
+
| Max seq len | 256 |
|
| 25 |
+
| Weight tying | True |
|
| 26 |
+
| Val loss | 3.3455 |
|
| 27 |
+
| Trained steps | 4900 |
|
| 28 |
+
|
| 29 |
+
## Training
|
| 30 |
+
|
| 31 |
+
Trained on **TinyStories + WikiText-2** with a custom BPE tokenizer.
|
| 32 |
+
|
| 33 |
+
### Hyperparameters
|
| 34 |
+
|
| 35 |
+
```json
|
| 36 |
+
{
|
| 37 |
+
"data_dir": "data/processed",
|
| 38 |
+
"save_dir": "runs/AtomSLM-600K",
|
| 39 |
+
"config": "AtomSLM-600K",
|
| 40 |
+
"steps": 5000,
|
| 41 |
+
"eval_every": 100,
|
| 42 |
+
"save_every": 500,
|
| 43 |
+
"batch_size": 32,
|
| 44 |
+
"seq_len": 256,
|
| 45 |
+
"lr": 0.0005,
|
| 46 |
+
"lr_min": 5e-05,
|
| 47 |
+
"warmup": 1000,
|
| 48 |
+
"grad_clip": 1.0,
|
| 49 |
+
"dropout": 0.1,
|
| 50 |
+
"device": "auto",
|
| 51 |
+
"resume": null,
|
| 52 |
+
"compile": false,
|
| 53 |
+
"amp": false,
|
| 54 |
+
"core_warmup_steps": 0
|
| 55 |
+
}
|
| 56 |
+
```
|
| 57 |
+
|
| 58 |
+
## Usage
|
| 59 |
+
|
| 60 |
+
```python
|
| 61 |
+
import torch
|
| 62 |
+
from models.atomgpt import AtomSLM
|
| 63 |
+
|
| 64 |
+
ckpt = torch.load('pytorch_model.bin', map_location='cpu')
|
| 65 |
+
model = AtomSLM(ckpt['config'])
|
| 66 |
+
model.load_state_dict(ckpt['model_state'])
|
| 67 |
+
model.eval()
|
| 68 |
+
```
|
| 69 |
+
|
| 70 |
+
## License
|
| 71 |
+
|
| 72 |
+
MIT
|
config.json
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"vocab_size": 4096,
|
| 3 |
+
"d_model": 120,
|
| 4 |
+
"n_layers": 8,
|
| 5 |
+
"ffn_mult": 2.0,
|
| 6 |
+
"dropout": 0.1,
|
| 7 |
+
"max_seq_len": 256,
|
| 8 |
+
"tie_weights": true,
|
| 9 |
+
"architecture": "AtomSLM",
|
| 10 |
+
"val_loss": 3.345488176345825,
|
| 11 |
+
"step": 4900
|
| 12 |
+
}
|
pytorch_model.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3eca566265982e026c15d839fe2692501945ff23d2dfb3d5adaa2d445424415f
|
| 3 |
+
size 2469423
|
tokenizer/tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|