Update README.md
Browse files
README.md
CHANGED
|
@@ -1,49 +1,81 @@
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
| 4 |
-
This file is intended strictly for saving the initial weights (checkpoint) of the JiRack GPT model. The model is "clean," meaning it contains no data and has never undergone pre-training.
|
| 5 |
|
| 6 |
-
|
| 7 |
|
| 8 |
-
|
|
|
|
| 9 |
|
| 10 |
-
|
| 11 |
|
| 12 |
-
|
|
|
|
|
|
|
| 13 |
|
| 14 |
-
|
| 15 |
|
| 16 |
-
|
| 17 |
|
| 18 |
-
|
| 19 |
|
| 20 |
-
|
|
|
|
| 21 |
|
| 22 |
-
|
| 23 |
-
MODEL_DIM = 768
|
| 24 |
-
NUM_HEADS = 12
|
| 25 |
-
NUM_LAYERS = 6
|
| 26 |
-
MAX_SEQ_LEN = 8192
|
| 27 |
-
FFN_HIDDEN_DIM = 4 * MODEL_DIM
|
| 28 |
-
HEAD_DIM = MODEL_DIM // NUM_HEADS
|
| 29 |
|
| 30 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
|
| 32 |
-
|
| 33 |
|
| 34 |
-
|
| 35 |
-
MODEL_DIM = 768
|
| 36 |
-
NUM_HEADS = 6
|
| 37 |
-
NUM_LAYERS = 6
|
| 38 |
-
MAX_SEQ_LEN = 8192
|
| 39 |
-
FFN_HIDDEN_DIM = 4 * MODEL_DIM
|
| 40 |
-
HEAD_DIM = MODEL_DIM // NUM_HEADS
|
| 41 |
|
| 42 |
-
|
| 43 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
|
|
|
|
|
|
|
| 45 |
|
|
|
|
| 46 |
|
|
|
|
| 47 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
|
|
|
|
|
|
| 49 |
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
|
|
|
| 4 |
|
| 5 |
+
# JiRack GPT Initial Weights
|
| 6 |
|
| 7 |
+
This file is strictly intended for saving the **initial weights (checkpoint)** of the JiRack GPT model.
|
| 8 |
+
The model is **"clean"**: it contains no data and has never undergone any pre-training.
|
| 9 |
|
| 10 |
+
It is engineered to be a maximally safe and robust base for **training from scratch** for specialized, smaller models, such as:
|
| 11 |
|
| 12 |
+
- **SPAM Detection Systems**
|
| 13 |
+
- **FRAUD Detection Models**
|
| 14 |
+
- **Background Check (BG Check) Models**
|
| 15 |
|
| 16 |
+
_A product of CMS Manhattan._
|
| 17 |
|
| 18 |
+
---
|
| 19 |
|
| 20 |
+
## Tokenizer Choices
|
| 21 |
|
| 22 |
+
- For English: **GPT-2 Hugging Face tokenizer**
|
| 23 |
+
- For multilingual use: **BERT tokenizer** from the Hugging Face library
|
| 24 |
|
| 25 |
+
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
| 27 |
+
## Model Architecture Details
|
| 28 |
+
|
| 29 |
+
### GPT-2 Architecture (Classic, Transformer-like)
|
| 30 |
+
|
| 31 |
+
```
|
| 32 |
+
CustomEmbedding
|
| 33 |
+
FrozenSignatureLayer
|
| 34 |
+
LearnedPositionalEmbedding
|
| 35 |
+
[TransformerBlock]
|
| 36 |
+
├── MultiHeadAttention
|
| 37 |
+
├── LayerNorm
|
| 38 |
+
├── LayerNorm
|
| 39 |
+
├── FFN
|
| 40 |
+
├── Linear
|
| 41 |
+
├── Activation: GELU
|
| 42 |
+
└── Linear
|
| 43 |
+
LayerNorm
|
| 44 |
+
Linear
|
| 45 |
+
```
|
| 46 |
|
| 47 |
+
---
|
| 48 |
|
| 49 |
+
## Model Checkpoint File Explanations
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
|
| 51 |
+
### **12-head Attention Model**
|
| 52 |
|
| 53 |
+
**Parameters:**
|
| 54 |
+
- `VOCAB_SIZE = 50257`
|
| 55 |
+
- `MODEL_DIM = 768`
|
| 56 |
+
- `NUM_HEADS = 12`
|
| 57 |
+
- `NUM_LAYERS = 6`
|
| 58 |
+
- `MAX_SEQ_LEN = 8192`
|
| 59 |
+
- `FFN_HIDDEN_DIM = 4 * MODEL_DIM`
|
| 60 |
+
- `HEAD_DIM = MODEL_DIM // NUM_HEADS`
|
| 61 |
|
| 62 |
+
**File:**
|
| 63 |
+
`JiRack_H12_L6_V50257_D768_MSL8192_FF768x4.pt`
|
| 64 |
|
| 65 |
+
---
|
| 66 |
|
| 67 |
+
### **6-head Attention Model**
|
| 68 |
|
| 69 |
+
**Parameters:**
|
| 70 |
+
- `VOCAB_SIZE = 50257`
|
| 71 |
+
- `MODEL_DIM = 768`
|
| 72 |
+
- `NUM_HEADS = 6`
|
| 73 |
+
- `NUM_LAYERS = 6`
|
| 74 |
+
- `MAX_SEQ_LEN = 8192`
|
| 75 |
+
- `FFN_HIDDEN_DIM = 4 * MODEL_DIM`
|
| 76 |
+
- `HEAD_DIM = MODEL_DIM // NUM_HEADS`
|
| 77 |
|
| 78 |
+
**File:**
|
| 79 |
+
`JiRack_H6_L6_V50257_D768_MSL8192_FF768x4.pt`
|
| 80 |
|
| 81 |
+
---
|