Update README.md
Browse files
README.md
CHANGED
|
@@ -2,29 +2,33 @@
|
|
| 2 |
language: en
|
| 3 |
license: mit
|
| 4 |
tags:
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
---
|
| 11 |
|
| 12 |
# TinyLM
|
| 13 |
|
| 14 |
-
A
|
| 15 |
|
| 16 |
## Architecture
|
| 17 |
|
| 18 |
| Hyperparameter | Value |
|
| 19 |
|---|---|
|
| 20 |
-
| Parameters |
|
| 21 |
| Layers | 4 |
|
| 22 |
| Hidden size | 64 |
|
| 23 |
| Attention heads | 4 |
|
| 24 |
| FFN dim | 192 |
|
| 25 |
| Embedding rank | 32 |
|
| 26 |
| Context length | 256 |
|
| 27 |
-
| Tokenizer | GPT-2 (
|
| 28 |
|
| 29 |
Uses a **factored (low-rank) embedding** to keep the vocab projection from eating the entire parameter budget, with weight tying on the output head.
|
| 30 |
|
|
@@ -36,7 +40,7 @@ Uses a **factored (low-rank) embedding** to keep the vocab projection from eatin
|
|
| 36 |
| Optimizer | AdamW (lr=3e-3, weight_decay=0.01) |
|
| 37 |
| Scheduler | Cosine annealing with warm restarts |
|
| 38 |
| Mixed precision | fp16 (torch.cuda.amp) |
|
| 39 |
-
| Hardware | Nvidia P100
|
| 40 |
|
| 41 |
## Usage
|
| 42 |
```python
|
|
@@ -44,10 +48,10 @@ from huggingface_hub import snapshot_download
|
|
| 44 |
import importlib.util
|
| 45 |
import torch
|
| 46 |
|
| 47 |
-
# Download
|
| 48 |
snapshot_download(repo_id="Fu01978/TinyLM", local_dir="./tinylm")
|
| 49 |
|
| 50 |
-
# Load via
|
| 51 |
spec = importlib.util.spec_from_file_location("modeling_tinylm", "./tinylm/modeling_tinylm.py")
|
| 52 |
module = importlib.util.module_from_spec(spec)
|
| 53 |
spec.loader.exec_module(module)
|
|
@@ -56,11 +60,6 @@ model, tokenizer, config = module.load_tinylm("./tinylm")
|
|
| 56 |
model.eval()
|
| 57 |
|
| 58 |
# Generate
|
| 59 |
-
output = module.generate(model, tokenizer, "Once upon a time")
|
| 60 |
print(output)
|
| 61 |
-
```
|
| 62 |
-
|
| 63 |
-
## Example Outputs
|
| 64 |
-
|
| 65 |
-
**Prompt:** Once upon a time
|
| 66 |
-
**Output:** Once upon a time there was a little girl named Mrs. She decided to go and be a little girl in the park. One day she had to go on a bed. From then on a lot of bread. She said, "What are you doing?" ...
|
|
|
|
| 2 |
language: en
|
| 3 |
license: mit
|
| 4 |
tags:
|
| 5 |
+
- tiny
|
| 6 |
+
- language-model
|
| 7 |
+
- causal-lm
|
| 8 |
+
- pytorch
|
| 9 |
+
datasets:
|
| 10 |
+
- roneneldan/TinyStories
|
| 11 |
+
- Skylion007/openwebtext
|
| 12 |
+
pipeline_tag: text-generation
|
| 13 |
+
library_name: transformers
|
| 14 |
---
|
| 15 |
|
| 16 |
# TinyLM
|
| 17 |
|
| 18 |
+
A 3.4M parameter causal language model trained from scratch, for experimentation.
|
| 19 |
|
| 20 |
## Architecture
|
| 21 |
|
| 22 |
| Hyperparameter | Value |
|
| 23 |
|---|---|
|
| 24 |
+
| Parameters | 3.403.968 |
|
| 25 |
| Layers | 4 |
|
| 26 |
| Hidden size | 64 |
|
| 27 |
| Attention heads | 4 |
|
| 28 |
| FFN dim | 192 |
|
| 29 |
| Embedding rank | 32 |
|
| 30 |
| Context length | 256 |
|
| 31 |
+
| Tokenizer | GPT-2 (50257 vocab) |
|
| 32 |
|
| 33 |
Uses a **factored (low-rank) embedding** to keep the vocab projection from eating the entire parameter budget, with weight tying on the output head.
|
| 34 |
|
|
|
|
| 40 |
| Optimizer | AdamW (lr=3e-3, weight_decay=0.01) |
|
| 41 |
| Scheduler | Cosine annealing with warm restarts |
|
| 42 |
| Mixed precision | fp16 (torch.cuda.amp) |
|
| 43 |
+
| Hardware | Nvidia P100 |
|
| 44 |
|
| 45 |
## Usage
|
| 46 |
```python
|
|
|
|
| 48 |
import importlib.util
|
| 49 |
import torch
|
| 50 |
|
| 51 |
+
# Download files
|
| 52 |
snapshot_download(repo_id="Fu01978/TinyLM", local_dir="./tinylm")
|
| 53 |
|
| 54 |
+
# Load via script
|
| 55 |
spec = importlib.util.spec_from_file_location("modeling_tinylm", "./tinylm/modeling_tinylm.py")
|
| 56 |
module = importlib.util.module_from_spec(spec)
|
| 57 |
spec.loader.exec_module(module)
|
|
|
|
| 60 |
model.eval()
|
| 61 |
|
| 62 |
# Generate
|
| 63 |
+
output = module.generate(model, tokenizer, "Once upon a time, ")
|
| 64 |
print(output)
|
| 65 |
+
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|