quantized with the default exl2 dataset with sequence lengths of 8192 and 400 calibration (stage 2, optimisation) lines instead of 2048/100. possibly microwaved, presumably better.

resulstant measurement file is present somewhere, though the default line count of 16 (still extended to 8192) was used for measurement (stage 1)

tokenizer works. tokenizer.model is not required for use with exllama2. no promises about sketchy software by "oobabooga"* :) try tabbyAPI/tavern, or exui if you don't miss CFG

consider yourselves lucky it's not a safetensors.zpaq this took all night to upload and YES i did refresh my access tokens after the Whoopsie, sorry!

*I'm sure it's fine it's just that I'll die if I ever see conda again.

DreamGen Opus V1

Models for (steerable) story-writing and role-playing.
All Opus V1 models, including quants.

Resources

Opus V1 prompting guide with many (interactive) examples and prompts that you can copy.
Google Colab for interactive role-play using opus-v1.2-7b.
Python code to format the prompt correctly.

story writing on dreamgen.com

Prompting

The models use an extended version of ChatML.

<|im_start|>system
(Story description in the right format here)
(Typically consists of plot description, style description and characters)<|im_end|>
<|im_start|>user
(Your instruction on how the story should continue)<|im_end|>
<|im_start|>text names= Alice
(Continuation of the story from the Alice character)<|im_end|>
<|im_start|>text
(Continuation of the story from no character in particular (pure narration))<|im_end|>
<|im_start|>user
(Your instruction on how the story should continue)<|im_end|>
<|im_start|>text names= Bob
(Continuation of the story from the Bob character)<|im_end|>

The Opus V1 extension is the addition of the text role, and the addition / modification of role names.

Pay attention to the following:

The text messages can (but don't have to have) names, names are used to indicate the "active" character during role-play.
There can be multiple subsequent message with a text role, especially if names are involved.
There can be multiple names attached to a message.
The format for names is names= {{name[0]}}; {{name[1]}}, beware of the spaces after names= and after the ;. This spacing leads to most natural tokenization for the names.

While the main goal for the models is great story-writing and role-playing performance, the models are also capable of several writing related tasks as well as general assistance.

Here's how you can prompt the model for the following tasks

Steerable Story-writing and Role-playing:
- Input:
  - System prompt: You provide story / role-play description, which consists of:
    - Plot description
    - Style description
    - Characters and their descriptions
  - Conversation turns:
    - Text / message turn: This represents part of the story or role play
    - Instruction: This tells the model what should happen next
- Output: Continuation of the story / role-play.
Story plot summarization
- Input: A story, or a few chapters of a story.
- Output: A description of the story or chapters.
Story character description
- Input: A story, or a few chapters of a story, set of characters.
- Output: A description of the characters.
Story style description
- Input: A story, or a few chapters of a story.
- Output: A description the style of the story.
Story description to chapters
- Input: A brief plot description and the desired number of chapters.
- Output: A description for each chapter.
And more...

Sampling params

For story-writing and role-play, I recommend "Min P" based sampling with min_p in the range [0.01, 0.1] and with temperature in the range [0.5, 1.5], depending on your preferences. A good starting point would be min_p=0.1; temperature=0.8.

You may also benefit from setting presence, frequency and repetition penalties, especially at lower temperatures.

Dataset

The fine-tuning dataset consisted of ~100M tokens of steerable story-writing, role-playing, writing-assistant and general-assistant examples. Each example was up to 31000 tokens long.

All story-writing and role-playing examples were based on human-written text.

Running the model

The model is should be compatible with any software that supports the base model, but beware of prompting and tokenization.

I recommend using these model versions:

7B: no quant (opus-v1.2-7b)
34B: no quant (opus-v1-34b) or awq (opus-v1-34b-awq)

Running on DreamGen.com (free)

You can try the model for free on dreamgen.com — note that an account is required.

Running Locally

Make sure your prompt is as close as possible to the Opus V1
- Regardless of which backend you use, it's important that you format your prompt well and that the tokenization works correctly.
- Read the prompt guide
- Read the prompt formatting code
- Make sure <|im_start|> and <|im_end|> are tokenized correctly
vLLM
- Google Colab: This is a simple interactive Google Colab to do role-play with the 7B model, it should fit on the T4 GPU.
- Code: This is simple script for interactive chat for one hard-coded scenario.
SillyTavern
- Settings, v2 kindly provided by @MarinaraSpaghetti
- Settings screenshot
- This is just an attempt at approximating the Opus V1 prompt, it won't be perfect
LM Studio
- Config
- Just like ChatML, just changed "assistant" to "text" role.
HuggingFace
- Chat template
- Just like ChatML, just changed "assistant" to "text" role.

Known Issues

34B tokenization:
- There seems to be a mismatch between the tokenizer of the base and fine-tuned model. It's unclear whether this also affected training, or whether it's just incorrectly saved tokenizer (you can see tokenizer.json was not saved (bug report)).
- This affects BOS and EOS (which aren't really used by Yi) and the tokenization of the first input token.
- Overall impact should be minor.
34B repetition:
- The 34B sometimes gets stuck repeating the same word, or synonyms. This seems to be a common problem across various Yi 34B fine-tunes.
GGUF:
- The conversion might be messed up and in my tests even Q_8 of the opus-v1.2-7b is much worse than the fp16 version.
Ooba:
- The tokenization might be messed up. Some users reported that <|im_start|> and <|im_end|> are tokenized as multiple tokens.

Community

Join the DreamGen community on Discord to get early access to new models.

License

This model is intended for personal use only, other use is not permitted.