Commit
•
cca0510
1
Parent(s):
831952d
Update model architecture
Browse files
README.md
CHANGED
@@ -65,7 +65,7 @@ for seq in sequences:
|
|
65 |
|
66 |
### Direct Use
|
67 |
|
68 |
-
Research on large language models,
|
69 |
|
70 |
### Out-of-Scope Use
|
71 |
|
@@ -127,13 +127,16 @@ Falcon-RW-1B was trained on 32 A100 40GB GPUs, using only data parallelism with
|
|
127 |
|
128 |
Hyperparameters were adapted from the GPT-3 paper ([Brown et al., 2020](https://arxiv.org/abs/2005.14165)).
|
129 |
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
-
|
|
|
|
|
135 |
|
136 |
-
|
|
|
137 |
|
138 |
Training happened in early December 2022 and took about six days.
|
139 |
|
@@ -149,6 +152,16 @@ Training happened in early December 2022 and took about six days.
|
|
149 |
|
150 |
Falcon-RW-1B is a causal decoder-only model trained on a causal language modeling task (i.e., predict the next token).
|
151 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
### Compute Infrastructure
|
153 |
|
154 |
#### Hardware
|
|
|
65 |
|
66 |
### Direct Use
|
67 |
|
68 |
+
Research on large language models, specifically the influence of adequately filtered and deduplicated web data on the properties of large language models (fairness, safety, limitations, capabilities, etc.).
|
69 |
|
70 |
### Out-of-Scope Use
|
71 |
|
|
|
127 |
|
128 |
Hyperparameters were adapted from the GPT-3 paper ([Brown et al., 2020](https://arxiv.org/abs/2005.14165)).
|
129 |
|
130 |
+
| **Hyperparameter** | **Value** | **Comment** |
|
131 |
+
|--------------------|------------|-------------------------------------------|
|
132 |
+
| Precision | `bfloat16` | |
|
133 |
+
| Optimizer | AdamW | |
|
134 |
+
| Learning rate | 2e-4 | 500M tokens warm-up, cosine decay to 2e-5 |
|
135 |
+
| Weight decay | 1e-1 | |
|
136 |
+
| Batch size | 512 | 4B tokens ramp-up |
|
137 |
|
138 |
+
|
139 |
+
#### Speeds, Sizes, Times
|
140 |
|
141 |
Training happened in early December 2022 and took about six days.
|
142 |
|
|
|
152 |
|
153 |
Falcon-RW-1B is a causal decoder-only model trained on a causal language modeling task (i.e., predict the next token).
|
154 |
|
155 |
+
The architecture is adapted from the GPT-3 paper ([Brown et al., 2020](https://arxiv.org/abs/2005.14165)), but uses ALiBi ([Ofir et al., 2021](https://arxiv.org/abs/2108.12409)) and FlashAttention ([Dao et al., 2022](https://arxiv.org/abs/2205.14135)).
|
156 |
+
|
157 |
+
| **Hyperparameter** | **Value** | **Comment** |
|
158 |
+
|--------------------|-----------|----------------------------------------|
|
159 |
+
| Layers | 24 | |
|
160 |
+
| `d_model` | 2048 | |
|
161 |
+
| `head_dim` | 64 | Reduced to optimise for FlashAttention |
|
162 |
+
| Vocabulary | 50304 | |
|
163 |
+
| Sequence length | 2048 | |
|
164 |
+
|
165 |
### Compute Infrastructure
|
166 |
|
167 |
#### Hardware
|