GPT-J-Skein - GGML Edition
This is the GGML port of our classic GPT-J-Skein model, a 6B model focussed on text adventures with additional novel data. It was a beloved text adventure and even writing model, back in the day people used the anti You bias userscript to enhance its writing ability. Later it was remade as Skein-20B which we also intend to convert to GGUF.
GGML in 2024, really?
Yes, GPT-J never saw adoption by Llamacpp and until this changes we have to rely on older code that originated from the pygmalioncpp project and that still lives on in KoboldCpp today. This model release was tested to work in KoboldCpp 1.66, but due to the age of the format does come with limitations.
What are the limitations of this conversion?
This format dates back to a time where K quants did not exist yet, so you will only be able to use regular quants or the FP16 version. Likewise a lot of modern features will be missing from the engine, you can still use smartcontext but you can't use context shifting. You can offload if you have a CUDA compatible GPU (ROCm is untested but may work), for full acceleration it is required to have every layer on the GPU. For non Nvidia GPU's you can use CLBlast to speedup the prompt processing, Vulkan does not support these older GGML models as it does not exist in our legacy code. Rope scaling even though its a much newer feature should be compatible, we also expect some of the more modern samplers to be compatible.
I don't use KoboldCpp, can I use it in X?
No, this upload is only meant for use with KoboldCpp. If you haven't tried KoboldCpp yet go give it a try! You can find it on https://koboldai.org/cpp
How was this conversion done?
Inside KoboldCpp's source code you can find otherarch/tools/convert_hf_gptj.py The relevant quantize_gptj can be compiled by using make tools in the KoboldCpp source code root directory.
Model Card for GPT-J-6B-Skein
Model Details
Model Description
- Developed by: KoboldAI
- Shared by [Optional]: KoboldAI
- Model type: Text Generation
- Language(s) (NLP): English
- License: Apache License 2.0
- Related Models: GPT-J 6B
- Parent Model: GPT-J
- Resources for more information:
Uses
Direct Use
This model is designed for creative story generation. It can understand both free-form text and text written in interactive fiction style with actions starting with "> You", such as:
You become aware of her breathing -- the slight expansion of her ribs, the soft exhalation -- natural, and yet somehow studied. "Ah -- by the way," she says, in a way that utterly fails to be casual, "have you seen the artist out there? -- My artist, that is."
"No," you respond, uneasy. You open your mouth and close it again.
> You ask about the experience of waking up
Downstream Use [Optional]
More information needed
Out-of-Scope Use
The model should not be used to intentionally create hostile or alienating environments for people.
Bias, Risks, and Limitations
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. 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 for a more detailed analysis of the biases in the Pile. 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.
See the GPT-J 6B model card for more information.
Recommendations
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
Training Details
Training Data
The data are mostly comprised of light novels from the dataset of the KoboldAI/GPT-Neo-2.7B-Horni-LN model and assorted interactive fiction. The dataset uses [Themes: <comma-separated list of genres>] for tagging, which means that if similar text is placed in the context, the model will attempt to generate text in the specified style(s). For more details about the dataset, consult this document.
Training Procedure
Preprocessing
The data were preprocessed using the Python package ftfy to eliminate as much as possible non-ASCII punctuation characters and possible encoding errors. The interactive fiction in the dataset also underwent deduplication since interactive fiction logs often contain duplicate text from, for example, visiting the same in-game area several times. spaCy was used for grammatical analysis with the purpose of reformatting the actions commonly found in old text adventure games into more complete sentences. There was also some manual elimination of things such as "thank you for playing" messages and title messages.
Speeds, Sizes, Times
Training took approximately 14 hours in total, with the average speed being 5265 tokens per second.
Evaluation
Testing Data, Factors & Metrics
Testing Data
More information needed
Factors
Metrics
More information needed
Results
More information needed
Model Examination
More information needed
Environmental Impact
Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).
- Hardware Type: More information needed
- Hours used: More information needed
- Cloud Provider: More information needed
- Compute Region: More information needed
- Carbon Emitted: More information needed
Technical Specifications [optional]
Model Architecture and Objective
More information needed
Compute Infrastructure
More information needed
Hardware
More information needed
Software
https://github.com/kingoflolz/mesh-transformer-jax
Citation
BibTeX:
@misc{mesh-transformer-jax,
author = {Wang, Ben},
title = {{Mesh-Transformer-JAX: Model-Parallel Implementation of Transformer Language Model with JAX}},
howpublished = {\url{https://github.com/kingoflolz/mesh-transformer-jax}},
year = 2021,
month = May
}
Glossary [optional]
More information needed
More Information [optional]
More information needed
Model Card Authors [optional]
KoboldAI in collaboration with Ezi Ozoani and the Hugging Face team
Model Card Contact
More information needed
How to Get Started with the Model
Use the code below to get started with the model.
Click to expand
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("KoboldAI/GPT-J-6B-Skein")
model = AutoModelForCausalLM.from_pretrained("KoboldAI/GPT-J-6B-Skein")