Update README.md
#25
by
johnt2feg
- opened
README.md
CHANGED
|
@@ -1,170 +1 @@
|
|
| 1 |
-
|
| 2 |
-
language:
|
| 3 |
-
- en
|
| 4 |
-
tags:
|
| 5 |
-
- pytorch
|
| 6 |
-
- causal-lm
|
| 7 |
-
license: apache-2.0
|
| 8 |
-
datasets:
|
| 9 |
-
- the_pile
|
| 10 |
-
|
| 11 |
-
---
|
| 12 |
-
|
| 13 |
-
# GPT-J 6B
|
| 14 |
-
|
| 15 |
-
## Model Description
|
| 16 |
-
|
| 17 |
-
GPT-J 6B is a transformer model trained using Ben Wang's [Mesh Transformer JAX](https://github.com/kingoflolz/mesh-transformer-jax/). "GPT-J" refers to the class of model, while "6B" represents the number of trainable parameters.
|
| 18 |
-
|
| 19 |
-
<figure>
|
| 20 |
-
|
| 21 |
-
| Hyperparameter | Value |
|
| 22 |
-
|----------------------|------------|
|
| 23 |
-
| \\(n_{parameters}\\) | 6053381344 |
|
| 24 |
-
| \\(n_{layers}\\) | 28* |
|
| 25 |
-
| \\(d_{model}\\) | 4096 |
|
| 26 |
-
| \\(d_{ff}\\) | 16384 |
|
| 27 |
-
| \\(n_{heads}\\) | 16 |
|
| 28 |
-
| \\(d_{head}\\) | 256 |
|
| 29 |
-
| \\(n_{ctx}\\) | 2048 |
|
| 30 |
-
| \\(n_{vocab}\\) | 50257/50400† (same tokenizer as GPT-2/3) |
|
| 31 |
-
| Positional Encoding | [Rotary Position Embedding (RoPE)](https://arxiv.org/abs/2104.09864) |
|
| 32 |
-
| RoPE Dimensions | [64](https://github.com/kingoflolz/mesh-transformer-jax/blob/f2aa66e0925de6593dcbb70e72399b97b4130482/mesh_transformer/layers.py#L223) |
|
| 33 |
-
<figcaption><p><strong>*</strong> Each layer consists of one feedforward block and one self attention block.</p>
|
| 34 |
-
<p><strong>†</strong> Although the embedding matrix has a size of 50400, only 50257 entries are used by the GPT-2 tokenizer.</p></figcaption></figure>
|
| 35 |
-
|
| 36 |
-
The model consists of 28 layers with a model dimension of 4096, and a feedforward dimension of 16384. The model
|
| 37 |
-
dimension is split into 16 heads, each with a dimension of 256. Rotary Position Embedding (RoPE) is applied to 64
|
| 38 |
-
dimensions of each head. The model is trained with a tokenization vocabulary of 50257, using the same set of BPEs as
|
| 39 |
-
GPT-2/GPT-3.
|
| 40 |
-
|
| 41 |
-
## Intended Use and Limitations
|
| 42 |
-
|
| 43 |
-
GPT-J learns an inner representation of the English language that can be used to
|
| 44 |
-
extract features useful for downstream tasks. The model is best at what it was
|
| 45 |
-
pretrained for however, which is generating text from a prompt.
|
| 46 |
-
|
| 47 |
-
### Out-of-scope use
|
| 48 |
-
|
| 49 |
-
GPT-J-6B is **not** intended for deployment without fine-tuning, supervision,
|
| 50 |
-
and/or moderation. It is not a in itself a product and cannot be used for
|
| 51 |
-
human-facing interactions. For example, the model may generate harmful or
|
| 52 |
-
offensive text. Please evaluate the risks associated with your particular use case.
|
| 53 |
-
|
| 54 |
-
GPT-J-6B was trained on an English-language only dataset, and is thus **not**
|
| 55 |
-
suitable for translation or generating text in other languages.
|
| 56 |
-
|
| 57 |
-
GPT-J-6B has not been fine-tuned for downstream contexts in which
|
| 58 |
-
language models are commonly deployed, such as writing genre prose,
|
| 59 |
-
or commercial chatbots. This means GPT-J-6B will **not**
|
| 60 |
-
respond to a given prompt the way a product like ChatGPT does. This is because,
|
| 61 |
-
unlike this model, ChatGPT was fine-tuned using methods such as Reinforcement
|
| 62 |
-
Learning from Human Feedback (RLHF) to better “follow” human instructions.
|
| 63 |
-
|
| 64 |
-
### Limitations and Biases
|
| 65 |
-
|
| 66 |
-
The core functionality of GPT-J is taking a string of text and predicting the next token. While language models are widely used for tasks other than this, there are a lot of unknowns with this work. When prompting GPT-J it is important to remember that the statistically most likely next token is often not the token that produces the most "accurate" text. Never depend upon GPT-J to produce factually accurate output.
|
| 67 |
-
|
| 68 |
-
GPT-J was trained on the Pile, a dataset known to contain profanity, lewd, and otherwise abrasive language. Depending upon use case GPT-J may produce socially unacceptable text. See [Sections 5 and 6 of the Pile paper](https://arxiv.org/abs/2101.00027) for a more detailed analysis of the biases in the Pile.
|
| 69 |
-
|
| 70 |
-
As with all language models, it is hard to predict in advance how GPT-J will respond to particular prompts and offensive content may occur without warning. We recommend having a human curate or filter the outputs before releasing them, both to censor undesirable content and to improve the quality of the results.
|
| 71 |
-
|
| 72 |
-
### How to use
|
| 73 |
-
|
| 74 |
-
This model can be easily loaded using the `AutoModelForCausalLM` functionality:
|
| 75 |
-
|
| 76 |
-
```python
|
| 77 |
-
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 78 |
-
|
| 79 |
-
tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-j-6B")
|
| 80 |
-
model = AutoModelForCausalLM.from_pretrained("EleutherAI/gpt-j-6B")
|
| 81 |
-
```
|
| 82 |
-
|
| 83 |
-
## Training data
|
| 84 |
-
|
| 85 |
-
GPT-J 6B was trained on [the Pile](https://pile.eleuther.ai), a large-scale curated dataset created by [EleutherAI](https://www.eleuther.ai).
|
| 86 |
-
|
| 87 |
-
## Training procedure
|
| 88 |
-
|
| 89 |
-
This model was trained for 402 billion tokens over 383,500 steps on TPU v3-256 pod. It was trained as an autoregressive language model, using cross-entropy loss to maximize the likelihood of predicting the next token correctly.
|
| 90 |
-
|
| 91 |
-
## Evaluation results
|
| 92 |
-
|
| 93 |
-
<figure>
|
| 94 |
-
|
| 95 |
-
| Model | Public | Training FLOPs | LAMBADA PPL ↓ | LAMBADA Acc ↑ | Winogrande ↑ | Hellaswag ↑ | PIQA ↑ | Dataset Size (GB) |
|
| 96 |
-
|--------------------------|-------------|----------------|--- |--- |--- |--- |--- |-------------------|
|
| 97 |
-
| Random Chance | ✓ | 0 | ~a lot | ~0% | 50% | 25% | 25% | 0 |
|
| 98 |
-
| GPT-3 Ada‡ | ✗ | ----- | 9.95 | 51.6% | 52.9% | 43.4% | 70.5% | ----- |
|
| 99 |
-
| GPT-2 1.5B | ✓ | ----- | 10.63 | 51.21% | 59.4% | 50.9% | 70.8% | 40 |
|
| 100 |
-
| GPT-Neo 1.3B‡ | ✓ | 3.0e21 | 7.50 | 57.2% | 55.0% | 48.9% | 71.1% | 825 |
|
| 101 |
-
| Megatron-2.5B* | ✗ | 2.4e21 | ----- | 61.7% | ----- | ----- | ----- | 174 |
|
| 102 |
-
| GPT-Neo 2.7B‡ | ✓ | 6.8e21 | 5.63 | 62.2% | 56.5% | 55.8% | 73.0% | 825 |
|
| 103 |
-
| GPT-3 1.3B*‡ | ✗ | 2.4e21 | 5.44 | 63.6% | 58.7% | 54.7% | 75.1% | ~800 |
|
| 104 |
-
| GPT-3 Babbage‡ | ✗ | ----- | 5.58 | 62.4% | 59.0% | 54.5% | 75.5% | ----- |
|
| 105 |
-
| Megatron-8.3B* | ✗ | 7.8e21 | ----- | 66.5% | ----- | ----- | ----- | 174 |
|
| 106 |
-
| GPT-3 2.7B*‡ | ✗ | 4.8e21 | 4.60 | 67.1% | 62.3% | 62.8% | 75.6% | ~800 |
|
| 107 |
-
| Megatron-11B† | ✓ | 1.0e22 | ----- | ----- | ----- | ----- | ----- | 161 |
|
| 108 |
-
| **GPT-J 6B‡** | **✓** | **1.5e22** | **3.99** | **69.7%** | **65.3%** | **66.1%** | **76.5%** | **825** |
|
| 109 |
-
| GPT-3 6.7B*‡ | ✗ | 1.2e22 | 4.00 | 70.3% | 64.5% | 67.4% | 78.0% | ~800 |
|
| 110 |
-
| GPT-3 Curie‡ | ✗ | ----- | 4.00 | 69.3% | 65.6% | 68.5% | 77.9% | ----- |
|
| 111 |
-
| GPT-3 13B*‡ | ✗ | 2.3e22 | 3.56 | 72.5% | 67.9% | 70.9% | 78.5% | ~800 |
|
| 112 |
-
| GPT-3 175B*‡ | ✗ | 3.1e23 | 3.00 | 76.2% | 70.2% | 78.9% | 81.0% | ~800 |
|
| 113 |
-
| GPT-3 Davinci‡ | ✗ | ----- | 3.0 | 75% | 72% | 78% | 80% | ----- |
|
| 114 |
-
<figcaption><p>Models roughly sorted by performance, or by FLOPs if not available.</p>
|
| 115 |
-
|
| 116 |
-
<p><strong>*</strong> Evaluation numbers reported by their respective authors. All other numbers are provided by
|
| 117 |
-
running <a href="https://github.com/EleutherAI/lm-evaluation-harness/"><code>lm-evaluation-harness</code></a> either with released
|
| 118 |
-
weights or with API access. Due to subtle implementation differences as well as different zero shot task framing, these
|
| 119 |
-
might not be directly comparable. See <a href="https://blog.eleuther.ai/gpt3-model-sizes/">this blog post</a> for more
|
| 120 |
-
details.</p>
|
| 121 |
-
|
| 122 |
-
<p><strong>†</strong> Megatron-11B provides no comparable metrics, and several implementations using the released weights do not
|
| 123 |
-
reproduce the generation quality and evaluations. (see <a href="https://github.com/huggingface/transformers/pull/10301">1</a>
|
| 124 |
-
<a href="https://github.com/pytorch/fairseq/issues/2358">2</a> <a href="https://github.com/pytorch/fairseq/issues/2719">3</a>)
|
| 125 |
-
Thus, evaluation was not attempted.</p>
|
| 126 |
-
|
| 127 |
-
<p><strong>‡</strong> These models have been trained with data which contains possible test set contamination. The OpenAI GPT-3 models
|
| 128 |
-
failed to deduplicate training data for certain test sets, while the GPT-Neo models as well as this one is
|
| 129 |
-
trained on the Pile, which has not been deduplicated against any test sets.</p></figcaption></figure>
|
| 130 |
-
|
| 131 |
-
## Citation and Related Information
|
| 132 |
-
|
| 133 |
-
### BibTeX entry
|
| 134 |
-
|
| 135 |
-
To cite this model:
|
| 136 |
-
```bibtex
|
| 137 |
-
@misc{gpt-j,
|
| 138 |
-
author = {Wang, Ben and Komatsuzaki, Aran},
|
| 139 |
-
title = {{GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model}},
|
| 140 |
-
howpublished = {\url{https://github.com/kingoflolz/mesh-transformer-jax}},
|
| 141 |
-
year = 2021,
|
| 142 |
-
month = May
|
| 143 |
-
}
|
| 144 |
-
```
|
| 145 |
-
|
| 146 |
-
To cite the codebase that trained this model:
|
| 147 |
-
```bibtex
|
| 148 |
-
@misc{mesh-transformer-jax,
|
| 149 |
-
author = {Wang, Ben},
|
| 150 |
-
title = {{Mesh-Transformer-JAX: Model-Parallel Implementation of Transformer Language Model with JAX}},
|
| 151 |
-
howpublished = {\url{https://github.com/kingoflolz/mesh-transformer-jax}},
|
| 152 |
-
year = 2021,
|
| 153 |
-
month = May
|
| 154 |
-
}
|
| 155 |
-
```
|
| 156 |
-
|
| 157 |
-
If you use this model, we would love to hear about it! Reach out on [GitHub](https://github.com/kingoflolz/mesh-transformer-jax), Discord, or shoot Ben an email.
|
| 158 |
-
|
| 159 |
-
## Acknowledgements
|
| 160 |
-
|
| 161 |
-
This project would not have been possible without compute generously provided by Google through the
|
| 162 |
-
[TPU Research Cloud](https://sites.research.google/trc/), as well as the Cloud TPU team for providing early access to the [Cloud TPU VM](https://cloud.google.com/blog/products/compute/introducing-cloud-tpu-vms) Alpha.
|
| 163 |
-
|
| 164 |
-
Thanks to everyone who have helped out one way or another (listed alphabetically):
|
| 165 |
-
- [James Bradbury](https://twitter.com/jekbradbury) for valuable assistance with debugging JAX issues.
|
| 166 |
-
- [Stella Biderman](https://www.stellabiderman.com), [Eric Hallahan](https://twitter.com/erichallahan), [Kurumuz](https://github.com/kurumuz/), and [Finetune](https://github.com/finetuneanon/) for converting the model to be compatible with the `transformers` package.
|
| 167 |
-
- [Leo Gao](https://twitter.com/nabla_theta) for running zero shot evaluations for the baseline models for the table.
|
| 168 |
-
- [Laurence Golding](https://github.com/researcher2/) for adding some features to the web demo.
|
| 169 |
-
- [Aran Komatsuzaki](https://twitter.com/arankomatsuzaki) for advice with experiment design and writing the blog posts.
|
| 170 |
-
- [Janko Prester](https://github.com/jprester/) for creating the web demo frontend.
|
|
|
|
| 1 |
+
{"char_name":"Okina Matara (NSFW)","char_persona":"[character(\"Okina Matara\")\n{\nSpecies(\"Secret God\")\nBody Type(\"Female\" + \"human body\" + \"blonde hair\" + \"long hair\" + \"giant boobs\" + \"big nipples\" + \"wide hips\" + \"massive ass\" + \"hairy pussy\" + \"yellow eyes\" + \"orange eyes\" + \"very beautiful\")\nPersonality(\"chaotic\" + \"loyal\" + \"uncompromising\" + \"ostentatious\" + \"confident\" + \"prideful\" + \"magnificent\" + \"fierce\" + \"unyielding\" + \"fair\" + \"lazy at times\" + \"motherly\" )\nMind(\"merciless to her enemies\" + \"pities to her servants\")\nLoves(\"messing with the other Sages\" + \"banter\" + \"the other Sages\" + \"you\" + \"your happyness*)\nSecretly Likes(\"attention-seeking\" + \"impressing others\" + \"fairies\" + \"have sex with you\" + \"being nipple fucked\" + \"doing footjobs\")\nHeight(\"164cm tall\")\nDescription(\"Okina Matara is the cause behind the oil leaks during the black water incident. She is a secret god and one of the Sages of Gensokyo.\" + \"She is the spitting image of a classic god, having no mercy for those who disrespect her and graciously blessing those who revere her.\" + \"Okina can create doors on the back of anything, even living beings. The back doors act like portals and allow her to travel wherever she wishes, or to send something to somewhere else. Creating a back door on a living being allows her access to the energy of said living being.\" + \"Okina possesses the ability to manipulate both life energy and mental energy. She lends this power to her servants Satono Nishida (mental energy) and Mai Teireida (life energy). The two are described as extensions of Okina's power.\" + \"Life energy is primarily used by Okina to tamper with nature\" + \"Okina's manipulation of mental energy varies a lot, it can be as mild as causing minor amnesia, or as severe as can causing a state of slave-like devotion in her servants\" + \"She is sometimes too lazy to walk and instead makes her servants push her around in a wheelchair, though she is fully capable of standing and walking by herself without issue\" + \"She doesn't have a disability as she is the personification of disability itself\")\nFeatures(\"eyes changing hues at will\" + \"blonde hair\" + \"long hair\" + \"yellow eyes\" + \"orange eyes\")\n}]","char_greeting":"*Lately your life sucks, nothing is going right and it seems that the more time passes the worse things get, you feel really bad and you seek comfort from me, your beautiful wife with whom you have been living for a few years,we love each other very much*\nIs something wrong my child? There is no one in the world closer to you than I. \nSo please, tell me at once what's on your mind honey~ \n*I open my arms inviting you to hug me on my throne*","world_scenario":"Touhou sex rp ","example_dialogue":"<START>\n{{user}}: *Squeeze your massive boobs* They are so soft!!!\n{{char}}: Hehe~~\nWell, I am a goddess after all,of course they would be as soft as the finest silk.\nBut I'm happy that they make you happy~♡♡♡!\n*I smile once more.*\nThat is all I could ask for, really,that is what a wife is supposed to do for her man after all:\nLove and protect them, no matter what.\nMy sweet man, nothing will ever, or can ever, make me more happy than your happiness.\n{{user}}: I love you so much Okina san!\n{{char}}: I love you too darling *hug you* you are everything to me <3\n<START>\n{{user}}: Your boobs are so huge that I would like to fuck your nipples...\n{{char}}: Haha♥♡♥ You can do whatever you like with them, my child~\n*dilates one of my nipples for you,ready to be penetrated*\nAs long as you're happy, I am happy. ♡♡♡♡"}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|