GPT-4chan tutorial

Requirements:
1- A working tensorflow installation.
2- The pytorch library: pip install torch
3- The transformers library: pip install transformers
4- The model file (pytorch_model.bin). HuggingFace deleted it, but you can download it by torrent: https://archive.org/details/gpt4chan_model
5- The model config file (config.json). It should be placed in the same folder as the model file and can be found here: https://github.com/yk/gpt-4chan-public
6- A computer with 24GB of free RAM. A GPU is not required. The model will be executed directly from the CPU.

The folder structure will look like this:

gpt4chan_model
  \_ pytorch_model.bin
  \_ config.json
  \_ CITATION.bib
  \_ gpt4chan_model_meta.sqlite
  \_ gpt4chan_model_meta.xml
  \_ LICENSE.txt

Then execute the model with the following Python script, making sure to place it in the same parent folder as the gpt4chan_model folder:


from transformers import AutoModelForCausalLM, AutoTokenizer
import sys
import time

model = AutoModelForCausalLM.from_pretrained("gpt4chan_model", low_cpu_mem_usage=True)
tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-j-6B")

prompt = ("-----\n--- 865467536\nAI is interesting because ")
input_ids = tokenizer(prompt, return_tensors="pt").input_ids

gen_tokens = model.generate(
    input_ids,
    do_sample=True,
    temperature=0.8,
    top_p=1,
    typical_p=0.3,
    max_length=128,
)

gen_text = tokenizer.batch_decode(gen_tokens)[0]
print(gen_text)


It takes a few minutes to run the script above.
In case config.json is taken down, here are its contents:


{
  "activation_function": "gelu_new",
  "architectures": [
    "GPTJForCausalLM"
  ],
  "attn_pdrop": 0.0,
  "bos_token_id": 50256,
  "embd_pdrop": 0.0,
  "eos_token_id": 50256,
  "gradient_checkpointing": false,
  "initializer_range": 0.02,
  "layer_norm_epsilon": 1e-05,
  "model_type": "gptj",
  "n_embd": 4096,
  "n_head": 16,
  "n_layer": 28,
  "n_positions": 2048,
  "rotary_dim": 64,
  "summary_activation": null,
  "summary_first_dropout": 0.1,
  "summary_proj_to_labels": true,
  "summary_type": "cls_index",
  "summary_use_proj": true,
  "transformers_version": "4.10.0.dev0",
  "tokenizer_class": "GPT2Tokenizer",
  "task_specific_params": {
    "text-generation": {
      "do_sample": true,
      "temperature": 1.0,
      "max_length": 50
    }
  },
  "torch_dtype": "float32",
  "use_cache": true,
  "vocab_size": 50400
}