NovelAI Research Tool - nrt

A golang based client with:

• Minimum Viable Product implementation of a NovelAI service API client covering:
  • /user/login - to obtain authentication bearer tokens.
  • /ai/generate - to submit context and receive responses back from the AI
• Iterative testing based on JSON configuration files.

Downloading

It is recommended that if you are a normal MacOS or Windows user that you download the latest release from https://github.com/wbrown/novelai-research-tool/releases.

• If you are a MacOS user on an Intel machine, you will want the darwin-amd64, otherwise if you are on an Apple Silicon machine, you will want darwin-arm64.
• The vast majority of Windows users will want the windows-amd64 build.

I have left Linux builds as an exercise for the user, as it is generally very easy to build on Linux.

Building

You will need the golang language tools on your machine.

• Windows: Download and install from here: https://golang.org/doc/install
• Mac: If do not use brew, you could download from https://golang.org/doc/install; otherwise brew install go
• Linux: Use your package maager to install go or golang. If you're using Linux, I presume you know how to do this.

Get a copy of the source code either by:

• Downloading the ZIP file: https://github.com/wbrown/novelai-research-tool/archive/refs/heads/main.zip
• Or git clone https://github.com/wbrown/novelai-research-tool.git -- if you want to keep up to date, it's strongly recommended that you install git and use that. If you do, you can just type git pull to get the latest source code updates.

Once you have the source code extracted, go into the the source directory, and do the following from the command line:

• go get -u
• go build nrt-cli/nrt.go

This will produce a binary nrt file.

Setup

The nrt tool uses environment variables to hold your NovelAI username and password. They are:

• NAI_USERNAME
• NAI_PASSWORD

Windows:

On Windows, type the following in at the command prompt:

setx NAI_USERNAME [email protected]
setx NAI_PASSWORD password

You will need to restart the command shell to load these settings.

MacOS/Linux:

• On MacOS, edit the .zshrc file in your home directory..
• On Linux, eidt the .profile file in your home directory.

Add the following lines:

export NAI_USERNAME="[email protected]"
export NAI_PASSWORD="password"

Either re-login, or restart your terminal, or type the above two lines directly into your shell prompt.

Running

There is a test file in tests/need_help.json that you can run, by invoking:

• Windows: nrt tests/need_help.json
• MacOS/Linux: ./nrt tests/need_help.json

This will generate multiple output files in tests after about 30 minutes, each containing 10 iterations of 10 generations each.

Scenario Support

You may optionally provide nrt with a .scenario file directly without writing a test specification .json file. You just can't control things like permutations, or setting variables.

• Windows: nrt tests/a_laboratory_assistant.scenario
• MacOS/Linux: ./nrt tests/a_laboratory_assistant.scenario

It will produce a .json and .txt output file in the same directory as the .scenario file.

Output Processing Tip

You can use an utility called jq to massage the output JSON into something that is readable. An example usage is:

jq --raw-output ".[]|\"\n*********************************\nOUTPUT \(.settings.prefix)\n\(.result)\"" nrt_outputfile.json

For example:

$ jq --raw-output ".[]|\"\n*********************************
  \nOUTPUT \(.settings.prefix)\n\(.result)\"" 
  white_samurai_output-6B-v3-style_hplovecraft-2021-07-10T153527-0400.json
*********************************
OUTPUT style_hplovecraft
There is silence for several seconds before anyone answers. Then I hear footsteps
approaching the door, and then the sound of sliding bolts being drawn back. In moments,
a young man opens the door, wearing only his yukata, or loose-fitting white robe. He has
black hair tied into a ponytail, and he appears nervous. "Can I help you?" he asks,
speaking English.
"Yes," I say, "I was wondering if you knew what kind of establishment this is."
...

You can also add things to the output, for the variables you are interested in, such as temperature or top_k.

jq --raw-output ".[]|\"\n*********************************\nOUTPUT \(.settings.prefix) TEMP: \(.settings.temperature) TOP_K: \(.settings.top_k)\n\(.result)\"" nrt_outputfile.json

For example:

$ jq --raw-output ".[]|\"\n*********************************\nOUTPUT \(.settings.prefix)
  TEMP: \(.settings.temperature) TOP_K: \(.settings.top_k)\n\(.result)\""
  white_samurai_output-6B-v3-style_hplovecraft-2021-07-10T153527-0400.json
*********************************
OUTPUT style_epic_fantasy TEMP: 0.55 TOP_K: 100
The sound of footsteps approaches quickly from behind me, and then stops abruptly
 My eyes widen as I turn around slowly. Standing before me, holding his sword pointed
 toward me, is a man wearing black hakama pants and a white jacket over them. He has a
 mask covering his face, except for his mouth, nose and eye holes. His hands are empty,
 though he carries a small leather pouch containing something wrapped in cloth.
 ...

Details of test.json

The nrt tool accepts a single filename as an argument, the .json file containing test parameters. They are more or less self-explanatory, but I will highlight some specific ones:

• scenario - nrt supports .scenario files that you can download from NovelAI!
• prompt_filename - where to get the prompts, this is a txt file for easy editing of prompts without having to escape like you would in JSON.
• prompt - the prompt if you'd rather put it in the JSON itself.
• memory - NovelAI memory section, as text.
• authors_note - NovelAI author's note section as text.
• output_prefix - where you want the JSON output from the generations to go.
• iterations - how many times to run the test, effectively.
• generations - how many times to take the output, concatenate, and re-feed back into the AI, like an user.
• parameters - contains NovelAI configuration parameters according to the API's specifications.

The sample need_help.json can be used as a template, along with a need_help.txt prompt file. There is also an example need_help-6B_v3-vanilla.json and need_help-6B_v3-vanilla.txt file that contains an example of what output nrt produces.

Permutation Specs

There is a special field of the .json file, permutations. It is an array of permutation specs, each permutation spec is a combinatorial set in and of itself.

Most of the fields under parameters have a list counterpart under permutations.

You may find the current list of variables that may be permuted upon (here)[https://github.com/wbrown/novelai-research-tool/blob/main/nrt.go#L29]

For example, a permutation spec that permutes on model model might look like:

{ "permutations": { "model": [ "2.7B", "6B-v4" ] } }

This will cause the nrt tool to perform iterations across both models.

If we add other fields to permute on, such as prefix or prompt_path, it becomes a combinatorial exercise -- the below permutations example will generate 62 combinations and tests to perform:

{ "permutations": [ {
    "prompt_path": [ "need_help.txt", "need_help_an_memory.txt" ],
    "model": [ "2.7B", "6B-v4" ],
    "prefix": [ "vanilla", "style_arthurconandoyle", "style_edgarallanpoe",
      "style_hplovecraft", "style_shridanlefanu", "style_julesverne",
      "theme_19thcenturyromance", "theme_actionarcheology",
      "theme_airships", "theme_ai", "theme_darkfantasy",
      "theme_dragons", "theme_egypt", "theme_generalfantasy",
      "theme_huntergatherer", "theme_magicacademy", "theme_libraries",
      "theme_mars", "theme_medieval", "theme_militaryscifi",
      "theme_naval", "theme_pirates", "theme_postapocalyptic",
      "theme_rats", "theme_romanceofthreekingdoms",
      "theme_superheroes", "inspiration_crabsnailandmonkey",
      "inspiration_mercantilewolfgirlromance", "inspiration_nervegear",
      "inspiration_thronewars", "inspiration_witchatlevelcap" ] } ] }

The above permutation set will test all the combinations of:

• prompts with and without author's note and memory.
• models 2.7B, 6B-v4
• all the current AI modules.

Another example is if we wanted to permute on the temperature value:

{ "permutations": [ {
  "temperature": [ 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8 ] } ] }

The above will cause 8 tests to be created, covering the temperature in 0.1 steps from 0.1 to 0.8.

Placeholder Support

nrt supports NovelAI placeholders both as a way to populate .scenario with default values, and as a way to permute on them.

The relevant key to use in both the specifications and the permutations section is placeholders:

An example for the specifications settings -- any text in prompt, memory, or authors_note will have placeholders evaluated and substituted:

  "authors_note": "[Period: ${Period}]\n[Year: 2020 AD]\n[Style: Stream of consciousness, intelligent, funny]",
  "placeholders": {
    "Period": "In the Past"
  },

An example of permutation on placeholders; the below example will cause four different permutations to be created:

  "authors_note": [
    "[Period: ${Period}]\n[Year: 2020 AD]\n[Style: Stream of consciousness, intelligent, funny]",
    "[Period: ${Period}; Year: 2020 AD; Style: stream of consciousness, intelligent, funny]"
  ], 
  "placeholders": [
        {
          "Period": "Modern Day"
        },
        {
          "Period": "Present Times"
        }
  ]

Configuration Notes

As of the writing of this section:

The prefix takes the following values and requires the model attribute to have the 6B-v4 or euterpe-v2 value:

• vanilla
• general_crossgenre
• inspiration_crabsnailandmonkey
• inspiration_mercantilewolfgirlromance
• inspiration_nervegear
• inspiration_thronewars
• inspiration_witchatlevelcap
• style_algernonblackwood
• style_arthurconandoyle
• style_edgarallanpoe
• style_hplovecraft
• style_julesverne
• style_shridanlefanu
• theme_19thcenturyromance
• theme_actionarcheology
• theme_ai
• theme_airships
• theme_childrens
• theme_christmas
• theme_darkfantasy
• theme_dragons
• theme_egypt
• theme_generalfantasy
• theme_history
• theme_horror
• theme_huntergatherer
• theme_litrpg
• theme_magicacademy
• theme_magiclibrary
• theme_mars
• theme_medieval
• theme_militaryscifi
• theme_naval
• theme_philosophy
• theme_pirates
• theme_poeticfantasy
• theme_postapocalyptic
• theme_rats
• theme_romanceofthreekingdoms
• theme_romanempire
• theme_spaceopera
• theme_superheroes
• theme_textadventure
• theme_valentines
• theme_vikings
• theme_westernromance
• utility_lorebookgenerator

The model parameter takes the following values:

• 2.7B
• 6B-v4
• euterpe-v2
• krake-v2

Client

A simple golang client for nrt that can be found in the novelai-research-tools/client folder. After nrt is set up properly, click the build.bat file to build the client. This will automatically build the client.

The main purpose of this client is to test the API, namely the logit_bias functionality.

Logit biases work by increasing the weights of a token, which increases the chance it has to be generated by the AI. If the AI wasn't going to ever generate the token before, this will not make it generate it, unless the bias is incredibly high. This helps the AI keep on track, as the AI will be more likely to generate words that are related to how you want the story to go.

A bias of around -8.0 is as effective of banning the token, and around 1.0 to 3.0 makes it much more likely that it'll appear.

You can find tokens with the tokenizer on the site, remember to include a space before the word, the majority of the time, the tokens are different. Words that are split up into multiple tokens are significantly more of a pain to deal with, but increasing the bias of the first part, then reducing it for every token afterwords helps.

The format for tokens is:

  "authors_note": [[token id, bias as a float],[token,float],[token,float]],

The example provided in the folder was designed to create long, run-on sentences by reducing the likelyhood that punctuation tokens will appear.

The client has four options right now, which are used by typing the word IN FULL CAPITALS as your input.

RETRY - Retries the last input, giving you a new output.

SAVE - Saves the current story to prompt.txt.

BACK - Goes back to the last input/output, there is a maximum of 65536 entries, it may crash after this point.

EDIT - Edits the last output, this is very wonky and clunky, so be warned.

NEXT - Gives a list of potential next tokens that could appear. This list represents the liklihood of tokens appearing BEFORE logit_bias is applied.

Along with this, I've added a grab_logits.exe file, editing the logits.txt file will let you grab the json entry for logit bias much faster. Enjoy!

Adventure Game

As a reward for reaching the very end of the document, there's a special treat. An Adventure module that replicates the classic Zork experience.

To build it, go into the adventure subdirectory and type go build adventure.go. This will produce a binary adventure or adventure.exe.

Run it by invoking ./adventure or adventure.exe, and you will be brought to a prompt:

You are in a maze of twisty passages, all alike. There are exits to the north, east, south, and west.
> 

Enjoy!