Update README.md
Browse files
README.md
CHANGED
|
@@ -8,6 +8,7 @@ tags:
|
|
| 8 |
- pytorch
|
| 9 |
- causal-lm
|
| 10 |
- code-generation
|
|
|
|
| 11 |
|
| 12 |
|
| 13 |
license: apache-2.0
|
|
@@ -29,55 +30,28 @@ This is a preliminary release of an experimental artifact and should be treated
|
|
| 29 |
|
| 30 |
|
| 31 |
| Hyperparameter | Value |
|
| 32 |
-
|
| 33 |
-
|
| 34 |
|----------------------|----------------------------------------------------------------------------------------------------------------------------------------|
|
| 35 |
-
|
| 36 |
-
|
| 37 |
| \\(n_{parameters}\\) | 1,331,810,304 |
|
| 38 |
-
|
| 39 |
-
|
| 40 |
| \\(n_{layers}\\) | 24 |
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
| \\(d_{model}\\) | 2,048 |
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
| \\(d_{ff}\\) | 8,192 |
|
| 47 |
-
|
| 48 |
-
|
| 49 |
| \\(n_{heads}\\) | 16 |
|
| 50 |
-
|
| 51 |
-
|
| 52 |
| \\(d_{head}\\) | 128 |
|
|
|
|
|
|
|
|
|
|
| 53 |
|
| 54 |
|
| 55 |
-
| \\(n_{ctx}\\) | 2,048 |
|
| 56 |
-
|
| 57 |
|
| 58 |
-
|
| 59 |
|
| 60 |
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
The model consists of 24 transformer layers with a model dimension of 2048, and a feedforward dimension of 8192. The model
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
dimension is split into 16 heads, each with a dimension of 128. Rotary Position Embedding (RoPE) is used.
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
The model is trained with the same tokenizer as GPT-NeoX-20b (link here), for a vocabulary of 50254 tokens.
|
| 76 |
|
| 77 |
|
| 78 |
## Training Data
|
| 79 |
|
| 80 |
-
The model was trained on the Pile, an 800Gb dataset composed of varied web corpora. The datasheet and paper for the Pile can be found [here] and [here] respectively
|
| 81 |
|
| 82 |
|
| 83 |
## Training Details
|
|
@@ -88,8 +62,7 @@ Following Bavarian et al. 2022, we train the model to additionally perform infil
|
|
| 88 |
|
| 89 |
Middle segments “to infill” were selected uniformly at random from contexts at the character level, and these contexts were then reformatted as
|
| 90 |
|
| 91 |
-
|
| 92 |
-
<SUF> {last 1/3rd of the context} <PRE> {first 1/3rd of the context} <MID> {middle 1/3rd of the context} <EOD>
|
| 93 |
|
| 94 |
|
| 95 |
|
|
@@ -118,11 +91,11 @@ model = AutoModelForCausalLM.from_pretrained("CarperAI/FIM-1.3b")
|
|
| 118 |
|
| 119 |
Suppose we have some text that we would like to perform infilling on at a certain “cursor location”.
|
| 120 |
|
| 121 |
-
This would have the form {some prelude text here} <INFILLING LOCATION> {some text following cursor}.
|
| 122 |
|
| 123 |
The way to perform infilling generation would be via placing the input text into this format:
|
| 124 |
|
| 125 |
-
<SUF> {some text following cursor} <PRE> {some prelude text here} <MID> ... language model output is generated after <MID> token!
|
| 126 |
|
| 127 |
|
| 128 |
## Intended Uses and Limitations
|
|
@@ -156,3 +129,5 @@ We also perform preliminary investigation on code generation and infilling capab
|
|
| 156 |
|
| 157 |
|
| 158 |
|
|
|
|
|
|
|
|
|
| 8 |
- pytorch
|
| 9 |
- causal-lm
|
| 10 |
- code-generation
|
| 11 |
+
- The Pile
|
| 12 |
|
| 13 |
|
| 14 |
license: apache-2.0
|
|
|
|
| 30 |
|
| 31 |
|
| 32 |
| Hyperparameter | Value |
|
|
|
|
|
|
|
| 33 |
|----------------------|----------------------------------------------------------------------------------------------------------------------------------------|
|
|
|
|
|
|
|
| 34 |
| \\(n_{parameters}\\) | 1,331,810,304 |
|
|
|
|
|
|
|
| 35 |
| \\(n_{layers}\\) | 24 |
|
| 36 |
+
| \\(d_{model}\\) | 2048 |
|
| 37 |
+
| \\(d_{ff}\\) | 8192 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
| \\(n_{heads}\\) | 16 |
|
|
|
|
|
|
|
| 39 |
| \\(d_{head}\\) | 128 |
|
| 40 |
+
| \\(n_{ctx}\\) | 2048 |
|
| 41 |
+
| \\(n_{vocab}\\) | 50254 |
|
| 42 |
+
| Positional Encoding | [Rotary Position Embedding (RoPE)](https://arxiv.org/abs/2104.09864)
|
| 43 |
|
| 44 |
|
|
|
|
|
|
|
| 45 |
|
| 46 |
+
The model consists of 24 transformer layers with a hidden dimension of 2048, and a feedforward intermediate dimension of 8192. The hidden dimension is split into 16 heads for self-attention, each with a dimension of 128. Rotary Position Embedding (RoPE) is used.
|
| 47 |
|
| 48 |
|
| 49 |
+
The model is trained with the same tokenizer as [GPT-NeoX-20b](https://arxiv.org/abs/2204.06745), for a vocabulary size of 50254 tokens.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
|
| 51 |
|
| 52 |
## Training Data
|
| 53 |
|
| 54 |
+
The model was trained on the Pile, an 800Gb dataset composed of varied web corpora. The datasheet and paper for the Pile can be found [here](https://arxiv.org/abs/2201.07311) and [here](https://arxiv.org/abs/2101.00027) respectively.
|
| 55 |
|
| 56 |
|
| 57 |
## Training Details
|
|
|
|
| 62 |
|
| 63 |
Middle segments “to infill” were selected uniformly at random from contexts at the character level, and these contexts were then reformatted as
|
| 64 |
|
| 65 |
+
\<SUF\> {last 1/3rd of the context} \<PRE\> {first 1/3rd of the context} \<MID\> {middle 1/3rd of the context} \<EOD\>
|
|
|
|
| 66 |
|
| 67 |
|
| 68 |
|
|
|
|
| 91 |
|
| 92 |
Suppose we have some text that we would like to perform infilling on at a certain “cursor location”.
|
| 93 |
|
| 94 |
+
This would have the form {some prelude text here} \<INFILLING LOCATION\> {some text following cursor}.
|
| 95 |
|
| 96 |
The way to perform infilling generation would be via placing the input text into this format:
|
| 97 |
|
| 98 |
+
\<SUF\> {some text following cursor} \<PRE\> {some prelude text here} \<MID\> ... language model output is generated after \<MID\> token!
|
| 99 |
|
| 100 |
|
| 101 |
## Intended Uses and Limitations
|
|
|
|
| 129 |
|
| 130 |
|
| 131 |
|
| 132 |
+
|
| 133 |
+
|