Single-Pass Language Model (SPLM)

A lightweight, zero-dependency Python chatbot that combines TF-IDF prompt matching, a dynamic local Markov chain, and a neural token predictor to generate conversational responses while requiring only one pass of training. It also includes a browser-based dark-mode interface for a cleaner chat experience.

---

Prerequisites

• Python 3.8+ (Uses standard library modules only—no pip install required).

---

1. Custom Data Setup

Create two plain text files in your project directory: prompts.txt and responses.txt. Each line in prompts.txt must correspond directly to the same line number in responses.txt. Two template files are provided.

prompts.txt

hello
how are you
what is your name
what do you do

responses.txt

hey there! how can I help you today?
i am doing great and ready to work.
my name is SPLM, nice to meet you.
i process text and generate contextual answers.

---

2. Training the Model

Run the train command to process your prompt/response pairs, build the TF-IDF vocabulary, and save the initialized model weights to a JSON file.

Basic Training

python3 splm.py train

Custom File Paths

python3 splm.py train --prompts my_prompts.txt --responses my_responses.txt --output my_model.json

Argument	Default	Description
--prompts	prompts.txt	Path to the prompt dataset file
--responses	responses.txt	Path to the response dataset file
--output	model.json	Path where the trained model JSON will be saved
--log-every	5	Progress print interval during indexing

---

3. Chatting with the Bot

Use the chat command to interact with your trained model.

Single Prompt Mode

python3 splm.py chat --prompt "hello how are you"

Interactive Terminal Mode Launch a continuous chat loop in your console:

python3 splm.py chat --interactive

(Press Return on an empty line to exit interactive mode.)

Advanced Options & Debugging

• Show Matched Prompts: Displays the top 5 TF-IDF matches and their similarity scores before generating an answer.

python3 splm.py chat --prompt "tell me a plan" --show-matches

• Enable Debugging Logs: Prints internal query vectors, top matches, and generation steps.

python3 splm.py chat --prompt "hello" --debug

• Limit Max Tokens: Control maximum response length.

python3 splm.py chat --interactive --max-tokens 20

Argument	Default	Description
--model	model.json	Path to the trained model JSON file
--prompt	"hello"	Input query (used in single-prompt mode)
--max-tokens	40	Maximum number of tokens to generate
--interactive	False	Launches interactive chat session
--show-matches	False	Prints top retrieved responses and TF-IDF scores
--debug	False	Prints verbose scoring and token prediction steps

---

4. Web Interface

Launch the browser UI with a local web server:

python3 web_chat.py --model model.json --open

The web interface uses a modern dark layout, message bubbles, a live input composer, and an optional match panel that shows the closest prompts used to build each reply.

SPLM (c) 2026 by npmInstallSnack