Documentation and deployment are essential in software development, though many can agree that these tasks are often tedious and time-consuming. This is where AutoDocs comes in!

AutoDocs uses a multi-model approach to automate the documentation or deployment process, using only an Asciinema recording of the code session.

Led by Julia Longtin (https://github.com/julialongtin) of the Human Feedback Foundation, AutoDocs provides deep analysis of a given code session in a hierarchical format, and is part of a greater initiative to and commitment towards creating a truly open-source LLM, where all datasets are ethically sourced (not stolen!).

Each model utilizes an 8B parameter distillation of Deepseek R1, fine-tuned using LoRA, which was trained on an A100 GPU provided by Google Colab.

1. Model 0 (Code Chunker): Splits code into <event> tags, representing unique actions within the code session.
2. Model 1 (Annotator): Summarizes and provides a "depth" value to each event provided by Model 0. A "depth" is represented as: -1 (subevent), 0 (independent event), >=1 (exiting events).
3. Model 3 (Dockerizer): A Dockerfile output of the commands run in the Asciinema recording.
4. Parsing Scripts: To avoid hallucination, outputs from models were further processed using parsers, avoiding complete regurgitation of input data. Currently, there exists Parser 0 (Asciinema to XML), Parser 1 (XML event grouping), and Parser 2 (XML summary insertion).

A raw Asciinema file is processed through a system (refer to System Architecture). All models can be loaded directly using the following code:

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("[model]")
model = AutoModelForCausalLM.from_pretrained("[model]")

A Python script that takes in a raw Asciinema file, and converts it into an XML. Edit each variable in the script to coincide with the desired directory and filenames:

    # Define the input and output directories:
    input_dir = "educational-AI-agent/data/Model 0/raw"
    output_dir = "educational-AI-agent/data/Model 0/formatted"

    # Name of the file to parse (must exist in input_dir).
    input_file_name = "renee_rec2.cast"
    # Name for the resulting XML output.
    output_file_name = "renee_rec2.cast.xml"

A fine-tuned model that will take in a converted XML (Parser 0 Output) and output a TXT with line numbers, printing "0" if the corresponding line should be the end of an event. Example:

1
2
3
0
4
5
6
7
0

Load the model directly with the earlier provided code snippet, available at: https://huggingface.co/bria7801/model-0

A Python script that takes a converted XML (Parser 0 Output) and a TXT with line numbers for code chunks (Model 0 Output). It will split relevant line numbers using <event> tags, with an <annotation> tag created before the closing <event> tag. On a Terminal or Command Shell application, run:

    python parser1.py <xml> <txt>

A fine-tuned model that will take an XML that is split by events (Parser 1 Output) and output a TXT with the hierarchical "depth" value and summary for each event. Example:

0
User connects to remote server via SSH using the command 'ssh 10.0.7.138' and enters password
-1
User changes their password on the remote server from '1M3T567!' to 'OpenYourHeartGPT'
1
User logs out of the SSH session and connection to the remote server is closed
0
User exits the terminal

Load the model directly with the earlier provided code snippet, available at: https://huggingface.co/bria7801/model-1

A Python script that will take an XML that is split by events (Parser 1 Output) and a TXT with hierarchical "depth" values and event summaries (Model 1 output), and insert the latter data into the former. Edit the plain_text_list and parsed_xml_list variables with the desired filepaths.

A fine-tuned model that will take an XML with summaries and depth values (Parser 2 Output) and output a Dockerfile that will run similar commands.

Load the model directly with the earlier provided code snippet, available at: https://huggingface.co/bria7801/model-3

1. Dataset Expansion: As all data is created by teams of Julia Longtin, and more data is needed to further increase the accuracy of the models.
2. Markdown Output: A model that is capable of outputting documentation, based on an annotated XML.
3. TinTin++ Output: A model that is capable of outputting TinTin++ code, based on the annotated XML.
4. Frontend: A frontend web app to improve user experience!

The current repository code has been modified and extended from the winter 2025 team's code base. Here's a detailed walkthrough of the winter 2025 team's code base: Walkthrough & Demo

For more information on the winter 2025 team's code base please see their white paper here: Winter 2025 White Paper

Most of the original team's code base is either outdated, or no longer working, hence it has been archived here: Archived Initial Code base