WPLM

WPLM

A small chat bot whose weights were written by hand as a plain-text edit list, with no training at all. You can follow each reply back to the neurons behind it. Needs internet access for a public lookup.

Made by Leejh123e on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: WPLM is a playful small-talk bot whose neural network was written by hand instead of trained. Its author set each nonzero weight through a plain-text list of edits. It skips gradient descent and any fitting to data. You can follow any reply back to the specific neurons, heads and weights responsible. The card says an AI wrote the generator script that produces the edit list.

Who it suits: It suits curious developers and researchers who want a small chat model they can read and edit like a text file, and teachers who explain how transformers work.

Industry | Who uses it | How they use it | Example
AI research | Interpretability researchers | Trace a reply to the weights that made it | Run chat.py with the verbose flag and read the lookups it printed
Education | Teachers of machine learning | Show that a transformer can exist without training | Open the edit file and follow one neuron line by line
Hobby projects | Tinkerers | Edit the weights as text and watch the bot change | Rename the bot by changing lines in the weight file

If you need a capable assistant, you can skip it. Its replies are playful and sometimes wrong.

Where to use it:

• Python on a laptop, since it needs only about 1 GB of RAM
• A computer with internet access, because word candidates come from a public lookup service
• Classrooms and demos about how transformers work
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Small talk | Anyone trying it | A short playful reply to a message
Traceability | Researchers | The lookups the network asked for, shown with the verbose flag
Weight editing | Tinkerers | A bot whose behavior changes when you change its edit list

How to set it up: What you need first:

• Python with numpy, sentencepiece and huggingface_hub
• Internet access, since the lookup service is public and the tokenizer is fetched on first use
• About 1 GB of free memory

1. Install the Python packages.
2. Download the repository into a folder named WPLM.
3. Go into that folder and run chat.py.

pip install -U huggingface_hub
hf download Leejh123e/WPLM

pip install numpy sentencepiece huggingface_hub
hf download Leejh123e/WPLM --local-dir WPLM
cd WPLM

How to use it: Run chat.py with a message to get one reply, or with no message to open an interactive chat. With the -v flag, it also shows each lookup the network requested.

python chat.py "tell me a joke"
python chat.py
python chat.py -v "hi"

You can also call it from Python.

import sys
sys.path.insert(0, "WPLM")
from wplm.load import load
from wplm.host import SimHost

bot = SimHost(load("WPLM"), seed=0)
text, info = bot.reply("what's your name?")
print(text)
print(info)

Tool | What it does
chat.py | Sends one message, or opens an interactive chat
wplm.load.load | Loads the frozen weights from the stored file
wplm.load.load_from_edits | Rebuilds the weights from the edit list, starting at zero
SimHost | Runs the lookups the network asks for and returns a reply

Example: You ask it to tell a joke. The weights write the opening, and the public lookup service supplies the ending. Ask who made it, and it answers from persona lines written into the weights, with no lookup at all.

For newcomers and experts: A good first try is asking for a joke with the -v flag, so you see what the network asked the lookup for. Then try a few other messages and compare.

For experts, the network is a decoder transformer with five blocks of attention plus a ReLU feed-forward part, 29 attention heads and no LayerNorm. The residual width is 9,114, and every dimension has a name in arch.json. Only 68,633 of its 249,124,749 parameters are nonzero. The weights live in an edit file of 14,396 lines, with a frozen copy stored as an npz file. It follows a dual language model idea. A statistical side offers next-word candidates with counts, and the programmed neural side chooses. That statistical side is the public infini-gram API over OLMo 2 data, built from 4.6 trillion tokens. The host program only tokenizes, runs lookups, supplies random numbers and appends tokens. For tokens it uses the stock Llama-2 SentencePiece model.

You send a message
The network decides what to ask the lookup service
It weighs the answer and picks a word or a line
The host adds each token until the network says stop

Good to know:

Title: Architecture at a glance
Part | Value
Attention heads | 29
Residual width | 9,114
Nonzero weights | 68,633
Total parameters | 249,124,749
Edit lines | 14,396
Training steps | 0

• Honesty: replies are playful and can be wrong. The card shows a wrong sum that came from trusting a bad match.
• Speed: the client holds itself to under 40 requests a minute, so each reply needs several seconds. Repeat lookups are saved in a cache folder in your home directory.
• Blocklist: the network is set to stop before it says certain blocked words.
• Authorship: the card says an AI wrote the script that generated the weights.
• License: mit
• Cost to run: no API key or paid service named.
• Data it reaches: it downloads files from Hugging Face, fetches a tokenizer on first use, and sends lookups based on your messages to a public service.
• Not for: serious questions, facts or maths.

On Hugging Face, 13 November 2026

• Task: text generation
• Downloads in the last 30 days: 0
• Likes: 5
• License: mit
• Languages: en
• Tags: weight-programming, weight-space-programming, programmed-weights, no-training, chatbot, infini-gram, text-generation

Links

• Model: https://huggingface.co/Leejh123e/WPLM
• Files: https://huggingface.co/Leejh123e/WPLM/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Leejh123e/WPLM

Then follow the usage in its model card. Nothing in this kit runs by itself.