Some questions

[gelsas]

Hi I just read about DOC and have some questions. I am developing in python since a few years but never really worked with AI yet. But this project is fascinating me because it may be the help/assistance I have been looking for.

Why am I interested so much in this ?
I love reading scifi books (I have read every big sci-fi book series there is and a lot of lesser known novels).
In the past few month I had a great idea for a scifi novel and have already a outline created for it. My biggest problem is that I am not a writer myself so I can not bring my idea to life or even create a more detailed outline of it.

Since I am a developer myself I already thought there must be something that can help me get around my shortcomings and after reading your paper just now, this sounds very promising.

My question now is can DOC be used to generate longer story's than 3,5k to 4k word long stories ?

I would love to use DOC to create and develop together with it a much longer rough story with me like 30000 to 40000 words long. I know it would probably not be perfect at all but it would be a great starting point for me to bring my idea for a story roughly to life and then just fine-tuning and fixing storyline issues my self. Then I could have a book based on my own idea. That would be awesome.

If writing such long stories is not possible, could DOC help me to create a very detailed outline a few thousand words long ?

If the answer is yes that would be possible then I will definitely test it myself but before investing all the time to understand every detail of DOC I thought it would make the most sense to ask here first.

Best,
Gelsas

[yangkevin2]

Hi Gelsas,

Thanks for reaching out! In principle, this repo is able to generate much longer stories such as the length you're suggesting, but the quality might be worse and you might lose a bit of coherence over time; I haven't tested stories of such length myself. It would also take a very long time to generate a story of that length through the public OPT API (3.5k-4k tokens already might already take a day or more, since each individual API call can take a few seconds), so if you don't have compute to set up your own OPT server you might want to just use GPT3 although the quality will be a bit lower. If I had to guess, the total cost of generating an entire 40k token story using our system through the GPT3 API (using the davinci model, and given an existing outline) would be on the order of $20.

Generating just the outline should also be possible (and in fact is recommended if you're trying to generate a much longer story). I haven't tested generating much more detailed outlines, but you can follow the instructions in the Plan + Outline Generation section of the README to generate the outline, setting --outline-levels to increase the depth of the outline. You could try a value like 5 or 6. There's a lot of filtering / reranking here so I'd guess it would cost $50 or so on the GPT3 API with default settings, but it can be made much cheaper as noted in options in the README.

Since you already have a story idea in mind, you could try to work with the outlining system "interactively" as well, although admittedly the procedure for doing so is a bit convoluted currently. For interactive outlining you'd use the --outline-restart-pkl option: start with a premise and generate using --outline-levels 1, then open the pkl and manually edit the outline object (the saved pickle object is a dict with an 'outline' key containing the outline, which you can print and then index into starting from the root node; then change the text, scene, and selected_entities at each node as needed). Then re-save the pkl and provide it to the plan generation and use --outline-levels 2, etc. (In principle, you could also work with the system interactively when you generate the main story after finishing the outline to help it stay on track, although we didn't add functionality for this yet. The easiest place to add it would probably be at https://github.com/yangkevin2/doc-story-generation/blob/main/scripts/main.py#L175, where you would manipulate beam[0].paragraphs_by_outline_section[outline_section] which will include the passages generated for the current outline item; make sure to also edit the last few passages in beam[0].all_paragraphs to maintain consistency.)

Hope that helps! To be totally upfront, we're not anywhere near the level of a strong human author yet; you can get a sense for average story quality by looking at the example outlines + stories at the end of the appendix in our paper. But if that level of writing looks acceptable to you for just initial outlining/drafting purposes, then go for it, and of course working with the system interactively will make the results a lot better. Please let me know if you have any other questions / difficulties.

Thanks,
Kevin

[gelsas]

Hi Kevin,

First of all I want to thank you for this detailed response. It already gives me a better understanding and guidance on what I need to do.

I am currently waiting to hear back from the Alpa guys on Slack but I might not be eligible to use the API, since I am not working for any institute but doing this research as a hobby. Do you know of any other ways to use OPT-175B via an API?

I don't think I would be eligible in Facebooks/Metas point of view since I am not someone researching at an institute. Also I am not sure how much compute I would need, do you have some more info regarding that ?

So currently it looks like I will need to go with Option 3: GPT3-175B (easy to get started; worse quality)

My questions here are the following:

Why is it not possible to use the detailed controller when using GTP3-175B?

Since OpenAI just released their chatgpt API: gpt-3.5-turbo (https://openai.com/blog/introducing-chatgpt-and-whisper-apis)
It is 10x cheaper and apparently better than the text-davinci-003 model.
Do you think this could be used instead of the current model that is used?

That would be all questions for know since I couldn't start my tests/experiments yet, but once I start testing I might have some more specific questions.

Just wanted to say thank you again for taking the time to respond to my questions.

Best,
gelsas

[yangkevin2]

Hi Gelsas,

I don't know of other OPT-175B APIs, unfortunately. You would also need a lot of compute to serve OPT-175B; recently they also released the LLaMA models https://ai.facebook.com/blog/large-language-model-llama-meta-ai/ which might be better and wouldn't require quite as much compute, but I'd guess you probably still need one or multiple high-end GPUs (40-80GBs) to serve it yourself if using the larger models, depending on which one exactly.

The reason we can't use the detailed controller with GPT3-175B is because the detailed controller relies on manipulating the logprobs of tokens at each step of generation. Not only does the GPT3 API severely limit how many logprobs you can get back, but it's unbelievably inefficient to use in this pipeline because it doesn't let you cache computation - you'd have to re-query with an entirely new prompt for every token you generate. The ChatGPT API will have the same issues. The Alpa API doesn't have these limitations (because I added the required functionality myself via a pull request a few months ago, lol). But it's not necessarily a dealbreaker if you can't use the detailed controller due to using GPT3 instead; just expect the model to wander off track away from your outline a bit more frequently. This can probably be remedied to a decent extent by manual intervention if you're willing to spend the effort to work with the system interactively paragraph by paragraph.

As for using ChatGPT or a text-davinci-00X model: during outline generation, due to generating the outline in breadth-first expansion order, you want the model to handle suffix context in addition to just a prefix (i.e., the prompt). Last I checked, text-davinci-003 and ChatGPT still didn't support this; we used text-davinci-002 in the paper. But if they do support suffix context you can swap them in.

For the main story generation where we use the base non-instruction-tuned davinci model, you also need to be a bit careful. My initial experiments with using text-davinci-002 instead were a complete disaster (it's mentioned in a footnote in our earlier Re3 paper https://arxiv.org/abs/2210.06774). I think the prompt we use, which contains hundreds of tokens of context, is totally out of distribution for these instruction-tuned models; in addition, the instruction-tuned models are usually finetuned to produce shorter outputs rather than full-length stories, whereas the base davinci's pretraining set actually contains a lot of novels which is what we want. We observed that text-davinci-002 usually generated maybe 1 paragraph of somewhat reasonable text before reliably starting to repeat itself or degenerating into gibberish. I think these problems are alleviated to some degree with text-davinci-003 and ChatGPT, but not completely gone: if you try to use these models I think you'll usually observe that they write in sort of a "higher-level" style compared to what you want in a story, i.e., they like to "tell" instead of "show."

It's definitely possible that you could get around these limitations with the newer text-davinci-003 and ChatGPT models with better prompt engineering, though, and it's worth exploring further. There are of course potential benefits. For example, text-davinci-003 and ChatGPT are definitely better at staying on-topic and not contradicting the context in the prompt compared to the base davinci model. So if you could resolve the style issues and use them as the inner model that DOC calls during story generation, it'd likely improve some of our metrics.

Hope that gives you some perspective on the possible choices.

Thanks,
Kevin

[gelsas]

Hi Kevin,

Thank you for the very detailed response.

I just got approved by meta for the LLaMA models and I am thinking of using vast.ai to rent the GPUs I probably need for this.
I have chatted with a few people and a few of them mentioned they used LLaMA-65B successfully on a single A100 80GB with 8bit quantization.

Using LLaMA is probably much better than using GPT3, correct?

Would it be possible to get some more insights on how I can use the LLaMA models with DOC ?

Can the detailed controller be used when using the LLaMA models ?

I hope you don't mind all these questions. Even if they are maybe "stupid" questions.
Pretty new to working with anything AI/ML related, basically just got into this after reading the DOC paper.

Best,
gelsas

[yangkevin2]

Hi Gelsas,

I haven't actually run the LLaMA models myself but I would guess it's better than GPT3 if you use the detail controller. There's no problem with using the detail controller in principle, but there could be some practical issues with tokenization though-- if the LLaMA models use different tokenization than OPT (I'm not sure if they do - you'd have to check) then you'd have to retrain the detail controller using a model that uses LLaMA tokenization. In this case you'd follow the instructions near the bottom of the readme to retrain it, and you'd probably also have to modify https://github.com/yangkevin2/doc-story-generation/blob/main/story_generation/common/controller/models/fudge_controller.py (see especially the bottom of the file where it has versions for GPT and OPT based models). Retraining might also be complicated since the smallest LLaMA model is 7B, which is a bit large for this purpose, so it would result in the detail controller running more slowly at inference time since it's not very performance-optimized currently. On the other hand, if it turns out that LLaMA uses the same tokenizer as OPT, then you can avoid all this trouble and directly use the existing detail controller I provide in the download.

Finally, you'd have to modify the code that interfaces with the 65B LLaMA model that you're calling, since right now it's assuming the Alpa interface for an OPT model.

TBH, I would consider starting with GPT3 anyway (you wouldn't need LLaMA for the planning stage anyway, only for the main drafting procedure after the planning is done) since it's so much easier to get started and try it out interactively, to get a rough sense of what the system can do. Then you could try to switch to LLaMA if you think it's worth the effort trying to improve it further for your purposes.

[gelsas]

Hi Kevin,

As per your suggestion I am starting with GPT3.

I have followed the instructions in the readme.
I ran into the following issue:

C:\Users\gelsas\Desktop\doc-story-generation-main>python -u scripts/main.py --controller none none none longformer_classifier --loader none none none order --controller-load-dir none none none doc_data/ckpt/outline_order_reranker --controller-model-string none none none roberta-large --no-editor --setup-only --outline-levels 3 --save-outline-file output/plan.pkl --log-file output/plan.log --premise PREMISE TEXT HERE --outline-levels 6
C:\Users\gelsas\AppData\Local\Programs\Python\Python310\lib\site-packages\torch\_masked\__init__.py:223: UserWarning: Failed to initialize NumPy: module compiled against API version 0xf but this version of numpy is 0xe (Triggered internally at  ..\torch\csrc\utils\tensor_numpy.cpp:68.)
  example_input = torch.tensor([[-3, -2, -1], [0, 1, 2]])
Traceback (most recent call last):
  File "C:\Users\gelsas\Desktop\doc-story-generation-main\scripts\main.py", line 7, in <module>
    from story_generation.edit_module.entity import *
  File "c:\users\gelsas\desktop\doc-story-generation-main\story_generation\edit_module\entity.py", line 15, in <module>
    from story_generation.common.util import *
  File "c:\users\gelsas\desktop\doc-story-generation-main\story_generation\common\util.py", line 25, in <module>
    from flair.data import Sentence
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python310\lib\site-packages\flair\__init__.py", line 25, in <module>
    from . import (  # noqa: E402 import after setting device
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python310\lib\site-packages\flair\models\__init__.py", line 1, in <module>
    from .clustering import ClusteringModel
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python310\lib\site-packages\flair\models\clustering.py", line 14, in <module>
    from flair.embeddings import DocumentEmbeddings
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python310\lib\site-packages\flair\embeddings\__init__.py", line 13, in <module>
    from .document import (
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python310\lib\site-packages\flair\embeddings\document.py", line 16, in <module>
    from flair.embeddings.token import FlairEmbeddings, StackedEmbeddings, TokenEmbeddings
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python310\lib\site-packages\flair\embeddings\token.py", line 10, in <module>
    import gensim
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python310\lib\site-packages\gensim\__init__.py", line 11, in <module>
    from gensim import parsing, corpora, matutils, interfaces, models, similarities, utils  # noqa:F401
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python310\lib\site-packages\gensim\corpora\__init__.py", line 6, in <module>
    from .indexedcorpus import IndexedCorpus  # noqa:F401 must appear before the other classes
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python310\lib\site-packages\gensim\corpora\indexedcorpus.py", line 14, in <module>
    from gensim import interfaces, utils
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python310\lib\site-packages\gensim\interfaces.py", line 19, in <module>
    from gensim import utils, matutils
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python310\lib\site-packages\gensim\matutils.py", line 1030, in <module>
    from gensim._matutils import logsumexp, mean_absolute_difference, dirichlet_expectation
  File "gensim\_matutils.pyx", line 1, in init gensim._matutils
ValueError: numpy.ndarray size changed, may indicate binary incompatibility. Expected 96 from C header, got 88 from PyObject

I am not sure why it shows this error:

UserWarning: Failed to initialize NumPy: module compiled against API version 0xf but this version of numpy is 0xe (Triggered internally at ..\torch\csrc\utils\tensor_numpy.cpp:68.)

Since as per the requirements.txt the version that got installed is: numpy==1.21.6

Any idea how this could be fixed?

[gelsas]

I was able to resolve this aswell. Switched from python 3.10 to 3.8.15

How long should the process approximately take ? To finish the oultine?

It was running for a bit and now it shows this in the plan.log: context length 1025 exceeded artificial context length limit 1024
and this in the command line: Token indices sequence length is longer than the specified maximum sequence length for this model (1025 > 1024). Running this sequence through the model will result in indexing errors

Not sure how to fix that, can you advice what is causing the context length to be exactly 1 character to long ?

This is the content of the plan.log

enerating characters and outline...
Initial Characters (will be filtered down later): 1.

Full Name: Taliyah Sutton

Character Portrait: Taliyah Sutton is a young woman with long brown hair and hazel eyes.

2.

Full Name: Jaxon Grey

Character Portrait: Jaxon Grey is a tall, muscular man with short black hair and blue eyes.

3.

Full Name: Nova Starr

Character Portrait: Nova Starr is a short, slender woman with long blonde hair and green eyes.

4.

Full Name: Zephyr Clarke

Character Portrait: Zephyr Clarke is a tall, lanky man with shortbrown hair and green eyes.

5.

Full Name: Riftmaster Arawn

Character Portrait: Riftmaster Arawn is a tall, imposing man with short white hair and blue eyes.

6.

Full Name: Overseer Kaelyn

Character Portrait: Overseer Kaelyn is a tall woman with long curly hair and blue eyes.

7.

Full Name: Druid Malachai

Character Portrait: Druid Malachai is a short man with long black hair and green eyes.

8.

Full Name: Senator Cassius

Character Portrait: Senator Cassius is a tall man with short grey hair and blue eyes.

9.

Full Name: Guardian Devon

Character Portrait: Guardian Devon is a tall woman with short blonde hair and green eyes.
Generating initial outline attempt 0
Expanding outline level 1
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 1
Expanding outline level 1
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 1
Expanding outline level 1
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 1
Expanding outline level 2
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 2
Expanding outline level 2
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 2
Expanding outline level 2
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 2
Expanding outline level 2
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 2
Expanding outline level 2
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 2
Expanding outline level 2
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 2
Expanding outline level 2
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 2
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Done expanding outline level 3
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Warning: didn't find acceptable outline
Expansion attempt 2 with temperature 1.4000000000000004
Warning: didn't find acceptable outline
Done expanding outline level 3
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 3
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 3
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 3
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Done expanding outline level 3
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Warning: didn't find acceptable outline
Expansion attempt 2 with temperature 1.4000000000000004
WARNING: no entities found, taking previous node's entities
Done expanding outline level 3
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Warning: didn't find acceptable outline
Expansion attempt 2 with temperature 1.4000000000000004
Done expanding outline level 3
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Done expanding outline level 3
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Warning: didn't find acceptable outline
Expansion attempt 2 with temperature 1.4000000000000004
Warning: didn't find acceptable outline
Done expanding outline level 3
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Warning: didn't find acceptable outline
Expansion attempt 2 with temperature 1.4000000000000004
Done expanding outline level 3
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Done expanding outline level 3
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Warning: didn't find acceptable outline
Expansion attempt 2 with temperature 1.4000000000000004
WARNING: no entities found, taking previous node's entities
Done expanding outline level 3
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Warning: didn't find acceptable outline
Expansion attempt 2 with temperature 1.4000000000000004
Warning: didn't find acceptable outline
Done expanding outline level 3
Expanding outline level 3
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Warning: didn't find acceptable outline
Expansion attempt 2 with temperature 1.4000000000000004
Warning: didn't find acceptable outline
Done expanding outline level 3
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Warning: didn't find acceptable outline
Expansion attempt 2 with temperature 1.4000000000000004
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Warning: didn't find acceptable outline
Expansion attempt 2 with temperature 1.4000000000000004
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: no entities found, taking previous node's entities
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
Warning: didn't find acceptable outline
Expansion attempt 2 with temperature 1.4000000000000004
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
Warning: didn't find acceptable outline
Expansion attempt 2 with temperature 1.4000000000000004
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
Done expanding outline level 4
Expanding outline level 4
Total nodes: 1
Expanding node 0
Expansion attempt 0 with temperature 1.2000000000000002
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
Warning: didn't find acceptable outline
Expansion attempt 1 with temperature 1.3000000000000003
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
WARNING: too many logit biases in outline character description generation, skipping uppercase biases
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

retrying...
context length 1025 exceeded artificial context length limit 1024

[yangkevin2]

Oh, now that I think about it, we never ran into this during our own tests since we limited the depth (number of outline levels) to 3 for our experiments. The problem is that we limit the GPT3/OPT context length for all methods to 1024 for fair comparison with baselines during our paper experiments, and you can end up exceeding that limit if you increase the depth further. You should be able to remove this artificial limitation by changing the --max-context-length argument, which defaults to 1024. The GPT3 API should support 2048 or 4096 depending on the model.

[gelsas]

I have added --max-context-length 2048 to my command but now I am getting this error:

Token indices sequence length is longer than the specified maximum sequence length for this model (1027 > 1024). Running this sequence through the model will result in indexing errors

And now a few hours later:

Token indices sequence length is longer than the specified maximum sequence length for this model (513 > 512). Running this sequence through the model will result in indexing errors
Traceback (most recent call last):
  File "c:\users\gelsas\desktop\doc-story-generation-main\story_generation\plan_module\plan.py", line 267, in generate_plan_info
    outline, (characters, character_strings, infer_attributes_string) = generate_outline(args, premise, setting, characters, character_strings, infer_attributes_string, instruct_model, outline_max_tokens, outline_levels=args.outline_levels, previous_outline=outline)
  File "c:\users\gelsas\desktop\doc-story-generation-main\story_generation\plan_module\plan.py", line 163, in generate_outline
    characters, character_strings, infer_attributes_string = node.recursively_expand_to_depth(
  File "c:\users\gelsas\desktop\doc-story-generation-main\story_generation\plan_module\outline.py", line 473, in recursively_expand_to_depth
    success, characters, character_strings, infer_attributes_string = node.expand(fixed_prefix,
  File "c:\users\gelsas\desktop\doc-story-generation-main\story_generation\plan_module\outline.py", line 400, in expand
    order_scores.append(get_outline_order_controller().evaluate_overall_texts([query]))
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\autograd\grad_mode.py", line 27, in decorate_context
    return func(*args, **kwargs)
  File "c:\users\gelsas\desktop\doc-story-generation-main\story_generation\common\controller\models\longformer_classifier.py", line 83, in evaluate_overall_texts
    outputs = self.model(**batch)
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\nn\modules\module.py", line 1194, in _call_impl
    return forward_call(*input, **kwargs)
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\transformers\models\roberta\modeling_roberta.py", line 1206, in forward
    outputs = self.roberta(
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\nn\modules\module.py", line 1194, in _call_impl
    return forward_call(*input, **kwargs)
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\transformers\models\roberta\modeling_roberta.py", line 841, in forward
    embedding_output = self.embeddings(
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\nn\modules\module.py", line 1194, in _call_impl
    return forward_call(*input, **kwargs)
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\transformers\models\roberta\modeling_roberta.py", line 133, in forward
    position_embeddings = self.position_embeddings(position_ids)
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\nn\modules\module.py", line 1194, in _call_impl
    return forward_call(*input, **kwargs)
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\nn\modules\sparse.py", line 160, in forward
    return F.embedding(
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\nn\functional.py", line 2210, in embedding
    return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
IndexError: index out of range in self

and a little bit after that I got this error:

Traceback (most recent call last):
  File "c:\users\gelsas\desktop\doc-story-generation-main\story_generation\plan_module\plan.py", line 267, in generate_plan_info
    outline, (characters, character_strings, infer_attributes_string) = generate_outline(args, premise, setting, characters, character_strings, infer_attributes_string, instruct_model, outline_max_tokens, outline_levels=args.outline_levels, previous_outline=outline)
  File "c:\users\gelsas\desktop\doc-story-generation-main\story_generation\plan_module\plan.py", line 163, in generate_outline
    characters, character_strings, infer_attributes_string = node.recursively_expand_to_depth(
  File "c:\users\gelsas\desktop\doc-story-generation-main\story_generation\plan_module\outline.py", line 473, in recursively_expand_to_depth
    success, characters, character_strings, infer_attributes_string = node.expand(fixed_prefix,
  File "c:\users\gelsas\desktop\doc-story-generation-main\story_generation\plan_module\outline.py", line 400, in expand
    order_scores.append(get_outline_order_controller().evaluate_overall_texts([query]))
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\autograd\grad_mode.py", line 27, in decorate_context
    return func(*args, **kwargs)
  File "c:\users\gelsas\desktop\doc-story-generation-main\story_generation\common\controller\models\longformer_classifier.py", line 83, in evaluate_overall_texts
    outputs = self.model(**batch)
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\nn\modules\module.py", line 1194, in _call_impl
    return forward_call(*input, **kwargs)
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\transformers\models\roberta\modeling_roberta.py", line 1206, in forward
    outputs = self.roberta(
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\nn\modules\module.py", line 1194, in _call_impl
    return forward_call(*input, **kwargs)
  File "C:\Users\gelsas\AppData\Local\Programs\Python\Python38\lib\site-packages\transformers\models\roberta\modeling_roberta.py", line 814, in forward
    buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
RuntimeError: The expanded size of the tensor (522) must match the existing size (514) at non-singleton dimension 1.  Target sizes: [1, 522].  Tensor sizes: [1, 514]

It has been running for 5-6 hours now. How long should it take ?

[yangkevin2]

The initial warning can be ignored -- it's complaining because we use the GPT2 tokenizer. GPT2 presumably has a max context length of up to 1024, but GPT3 shares the same tokenizer while using a longer context length.

I see in your initial command from earlier that you set --outline-levels to 6, which is probably excessive; I don't think I'd go more than 4 or 5. The outline itself is going to be growing exponentially in size as you increase the number of levels, and the time will increase correspondingly as well. You could run it with --outline-levels set to 3 first to verify that it works, then try 4, then maybe 5 if you still want it to be longer? Using 3 levels of outline usually takes several minutes, so 4 should take under an hour (probably not particularly close). 5 might take a couple of hours though if I had to guess. I'm not too surprised that you hit 5-6 hours with 6 levels of outline. I'm guessing that in your most recent run, it probably finished generating the first 5 levels of the outline and crashed while doing the 6th (you could check the logs if you saved them)

Your most recent crash looks like it's due to the context length getting too long for outline order reranker. The context length was fine for the length that we used in our paper experiments, but the outline is getting too long for it with your current settings. The problem might just go away if you decrease --outline-levels to 4 or 5, but otherwise you might need to retrain the outline order reranker described at the bottom of the readme using a model with longer context length (e.g. longformer instead of roberta); alternatively, truncate the context being fed into it as input at https://github.com/yangkevin2/doc-story-generation/blob/main/story_generation/plan_module/outline.py#L400 .

Sorry for all the trouble; we haven't carefully tested larger values of --outline-levels so it's somewhat new ground for me as well.