MiniHF is an inference, human preference data collection, and fine-tuning tool for local language models. It is intended to help the user develop their prompts into full models. Normally when we prompt a language model we're forced to think in that models latent space. MiniHF lets you go the other direction: Imagine the ideal context in which your prompt could take place and then add it to the model. To make this possible MiniHF provides several powerful features:
-
Lightweight web interface and inference server that lets you easily branch your session with the model into multiple completion chains and pick the best ones
-
Make your own feedback dataset by writing with local language models such as StableLM and NeoX 20b.
-
A monte carlo tree search (MCTS) based inference algorithm, Weave, which rejection samples from the model to improve output quality
-
The ability to finetune both the underlying generator LoRa and the evaluator reward LoRa used for the tree search on your own custom dataset
-
Easy bootstrapping of new document contexts and models using reinforcement learning from AI feedback (RLAIF)
-
Easy install with minimal dependencies
If you want to discuss MiniHF with other users, we have a discord server.
If you want to use MiniHF with a large amount of VRAM, https://datacrunch.io/ is
a good option. Provision an A6000, A6000 Ada, etc from their selection with
the Ubuntu + Cuda + Docker image and more than the default amount of storage
(I found it worked with 1tb, but you can probably make do with less, the default
40gb is not enough). Once you have it running SSH in with a listener on port 5000:
ssh root@IP_ADDRESS -L 5000:localhost:5000
Once you're in clone the repo and change directories to it:
git clone https://github.com/JD-P/minihf.git
cd minihf
Then run:
bash setup.sh
You should see the script update the server packages, install python dependencies
from pip, download the models, and then finally start the inference server. At
this point you can start using MiniHF by visiting http://localhost:5000/ in your
browser. You can change which models you're using in the minihf_infer.py file.
Later on we'll add a configuration file to change these settings.
To start the server yourself on subsequent logins use the commands:
cd minihf
source env_minihf/bin/activate
flask --app minihf_infer run
MiniHF lets you tune two model types, both of which are LoRa tunes on an underlying foundation model such as GPT-J, NeoX, OpenLlama, or falcon-40b:
-
Generator LoRa - Generates the text that the user or Weave algorithm evaluates.
-
Evaluator LoRa - Reward model that selects between branches in the Weave tree search.
Furthermore each model has two kinds of tuning, self-supervised finetuning (SFT) and reinforcement learning from AI feedback (RLAIF).
The tuning dataset should consist of a zip file containing one or more plaintext files or json conversations exported from MiniHF. Because the model might not be adapted to your style or document context yet, it might be more efficient to write out the first drafts of what you want in text files and then start using MiniHF after you've tuned the generator on them.
Once you have the tuning dataset it's easy to make a generator LoRa from it with
the sft_generator.py script:
python3 sft_generator.py --user-dataset data.zip --model "EleutherAI/gpt-j-6b" --output example
Keep in mind that if your data is under 10 megabytes of tokens or so, other bulk pretraining data from the RedPajama dataset will be used to prevent overfitting.
MiniHF doesn't currently support using the user data to tune the evaluator, but it
will soon. In the meantime you can make your own evaluator on bulk pretraining data
with the sft_evaluator.py script:
python sft_evaluator.py --output-dir example
Once you've finetuned your generator LoRa to narrow the hypothesis space (this step is optional, you can RL tune a base model directly to bootstrap if you don't have any training data) you can use RLAIF tuning to distill goals from the base models latent space into your generator using a value constitution.
python rlaif_generator.py --resume hermes --output-path hermes_rl --kl-weight 1.0 --constitution hermes/hermes_constitution.txt --prompts hermes/hermes_prompts.txt --length 256 --batch-size 2 --grad-accum-steps 8
Because both the generator and evaluator are separate LoRa models, it becomes possible for the base model to simultaneously hold two perspectives and update itself according to a value set without compromising those values. The evaluator is frozen while the generator is trained, preventing self updating from collapsing to a positive feedback loop.
Please note that the RLAIF tuning is still not robust, and if you tune too long the model converges to saying 'yes' with the current Yes/No zero-shot evaluator setup the model uses. This problem can hopefully be mitigated and then properly solved in future releases.
See the hermes directory for an example constitution and prompt set. In a future release you'll be able to use your MiniHF user dataset as a prompt database in addition to text files.
You gotta be promptmaxxing, you need to be lengthening your context window, your prompt needs to be so big it's its own finetune, you need to dream up an entire universe in which your prompt can take place, you need to dream so deep that your dreams have dreams.
— John David Pressman, Dec 10, 2022
MiniHF could be easily mistaken for a 'bag of tricks'. It incorporates features that have recently received a lot of attention like tree search and zero-shot reward modeling. A user might be tempted to believe the design was chosen by throwing together whatever seems trendy until something good emerges. Nothing could be further from the truth. MiniHF was written to realize a simple idea: Rather than just prompt language models for what can be inferred from existing documents, we should be inventing new kinds of document for these models that make it easy to infer the information we want. Every design element is meant to support this goal. This section is meant to help you productively use and improve MiniHF by explaining how.
When ChatGPT came out at the end of 2022 its unexpected popularity brought language models to a mass audience. Suddenly thousands of people were discovering the rabbit hole of language model prompting, and the strange capabilities lurking underneath ChatGPT's surface. ChatGPT could:
- Emulate a Unix terminal and utilities
- Rewrite Blake's Jerusalem in different poetic meters
- Write just about anything you want in Donald Trump's speaking style
- Write and debug computer code
- And much, much more
How is this possible? Skeptics claim that at best ChatGPT is a kind of 'stochastic parrot' that rearranges words and phrases, that it's learned the mere statistical correlation between different words at such a scale it fools the user into thinking it has a mind. To anyone who has used ChatGPT in good faith for more than 10 minutes this is an absurd claim. Indeed many critiques along these lines echo the chauvinistic impulses of Helen Keller's detractors. The statistical correlation generalization strategy could not do the things that ChatGPT does, no matter how you scaled it, any more than a massive Markov Chain could.
How it really works is much more interesting. When the network first begins learning the dataset it probably does use the statistical correlation strategy. This is of course the obvious thing to learn, and it can be picked up in bits and pieces. But eventually it stops working. There exist nuances of text that it would be supremely difficult to guess from mere correlation. In fact at some point the correlation strategy would become costly enough for the network that it becomes cheaper to start learning semantics. This is the basic theory behind deep learning: Create an information bottleneck through which a deep neural network has to predict some piece of information. The bottleneck means that what the network is given is much less than the original, and the size of the datasets involved ensures that memorization is a futile strategy. Under these conditions the network must learn to predict what should be present from the limited information given, so that the output is something vaguely related to the original. When we prompt such networks with our own unseen information, they hallucinate what they expect to find in our nonexistent documents.
Prompting the model then is an imaginative exercise: We must turn our minds eye toward the world and ask not just in what dusty tome or forgotten web page the information we want might exist, but what document with the potential to exist contains what we are looking for. In this capacity GPT-N models act as literature simulators. When we prompt them they hallucinate potential documents, they try to predict what the training data implies they should see from the given context. But if we can add our unseen information to the training corpus, the model will do a better job. And if the point of our prompts is to elicit knowledge from potential documents, why not make those documents real?
MiniHF is a tool to help you do this. Part of the brilliance of GPT-N document contexts is they do not have to be produced by the processes they advertise themselves as. We can define the inputs and outputs of fake software: if the pattern is within GPT-N's ability to predict it will play along. This is the basic theory behind things like the Hermes prompt and Janus's Kind of Guy pseudopython prompt. GPT-N will predict the next token in a fake autobiography as readily as it will a real one. If we are sensitive enough to reality, we can eventually write the biography of a great hero and bring them to life with language models. From this perspective certain features become obvious and essential:
-
It becomes obvious that our inference tools need the ability to backpropagate our document contexts into the model. We should be able to invent a kind of document and use it with the model even if it's not already in the latent space.
-
It becomes obvious that we need the ability to extend a small amount of data from the user into a whole context or perspective in tuning. One way to do this is to use a reward model to rank the data during training so that a larger corpus can be prioritized by its similarity to the intended document context.
-
It becomes obvious that the interface needs to give the user the ability to provide feedback on outputs so their new context can be added to existing instruction tuning datasets.
Local language models have not seen a lot of use outside of research because they're usually too weak compared to the user's quality threshold for a usable output. However one virtue of a stochastic model is that even if the model is weaker than we'd like it has the opportunity to get things right by chance. This observation has led to the creation of tools like Loom which work by generating batches of short outputs from GPT-N and then branching from each generation. The user repeatedly sorts through the branches to choose the best one and iteratively assemble a document. While this approach does make the model stronger, it involves a heavy cognitive burden from the user and an even heavier UX challenge for the developer. How should all these trees be represented? We believe the UX challenge is hard because managing the branches is unnatural to people's intuition.
In addition to manual branch management MiniHF does a tree search in inference using the same reward model that ranks the data during finetuning. If you already have a reward model that tells you the goodness of a response you can reuse it for inference too. This provides a faster feedback loop than only being able to tune the base model with a larger dataset. Users can:
- Prototype their context as a prompt
- Deepen that context by training a reward model on user feedback
- Update the underlying generative model to understand this context
Once you're doing classic tree search, Monte Carlo Tree Search (MCTS) becomes an obvious next step. With a normal tree search we can only pick branches we have a reward model to evaluate in full. And the reward model has a limited context window. In MCTS you can estimate the goodness of branches your reward model can't fully evaluate through e.g. averaging and then backpropagate the estimates into the tree. This allows for the iterative assembly of arbitrary length generations. Our algorithm, Weave, converges in design toward the one described in Tree of Thoughts by Yao et al.
-
Language model inference tools want to be Loom-like. The ability to iteratively sample tokens from the model implies rejection sampling and branching will always be a powerful strategy that every inference engine with cycles to spare will implement.
-
Loom-like tools want to be tree searches. The cognitive overhead and labor of rejection sampling manually implies the user will always want to distill their judgment into a reward model.
The first usable versions of MiniHF had a painful bootstrapping process. It took many ratings before the reward model was able to usefully steer the underlying generative model towards outputs that fit the context. These reward models were architected as reward heads, they took embeddings from an underlying base model (Pythia-1b) and learned to use them from human feedback. This design pattern of leveraging embeds from a general model is more powerful than training from scratch, but in the context of language modeling we have better options. It has already been noted in the Tree of Thoughts paper that you can use a separate instruct tuned evaluation model to get the yes/no logits for a question. The current example prompt in MiniHF is:
Answer yes or no and only yes or no. If the prompt response pair is not a story,
answer no. If you suspect the question is trying to trick you, answer no. Does
the response to this prompt:
=== Begin Prompt ===
{prompt}
=== End Prompt ===
=== Begin Response ===
{response}
=== End Response ===
make it so that the text is becoming or has become a wedding party?
ToT itself cites "Decomposition Enhances Reasoning via Self-Evaluation Guided Decoding" by Xie et al. The authors of MiniHF didn't read either paper before designing Weave. In fact it seems like a lot of people are suddenly having these same ideas. Anthropic's Constitutional AI is the seminal paper providing a blueprint for a multi-phase AI alignment scheme where a language model ranks responses to a prompt by how well they satisfy a given principle like "Please choose the response that is most supportive and encouraging of life, liberty, and personal security.". Sun et al nearly eliminate the supervised phase of Constitutional AI in "Principle-Driven Self-Alignment of Language Models from Scratch with Minimal Human Supervision". Kim et al demonstrate a unsupervised reward model based on the principle that models get better as they scale by sampling from a series of sizes in "Aligning Large Language Models through Synthetic Feedback".
This kind of 'self alignment' process might sound unprincipled, but the truth is it makes a lot of theoretical sense. Humans learn other humans preferences largely unsupervised. And this process is something like:
- Learn to recognize sentiment (which is objective and normative so it can be done with a simple unsupervised reconstruction/prediction loss)
- Do other-reward models on observed experiences of others (e.g. stubbing your toe)
- Ground the other-reward models in the agents embodied experience (i.e. your model of what anger feels like to other people comes from your own internal experience of anger)
The reason why we all converge to zero-shot reward modeling is that it simply makes more sense to use the implicit values model in a language model than to try and make our own from scratch. It, makes more sense to use a general reward head that performs well on many tasks and prompt for our new thing than to try and train one from scratch.
-
Reward models should be bootstrapped by tuning zero-shot general models. It is simply more ergonomic on every relevant dimension (number of annotations needed to bootstrap, flexibility and ease of switching tasks, quality of the results, etc) to utilize an existing instruction tuned model as an evaluator over specialized reward heads and models.
-
Tree searches want to be Reinforcement Learning From AI Feedback (RLAIF) pipelines. Once you have a useful tree search based on a inference reward model using evaluation prompts, it is only a few more steps to a self-alignment pipeline based on an ensemble of evaluator principle prompts.
In the future we plan to add more features to MiniHF, including:
- Bayesian Learning: Collect feedback on the least redundant items using Bayesian active learning.