L03 · Inside Heretic

L03 · Interactive walkthrough

Inside Heretic

Follow the inputs, find a direction, change the weights, inspect the result.

Explore the workflow

A companion to L03-easy-heretic.ipynb. This page explains the process; model execution happens in Python. Answer panels replay a recorded lab run.

Notebook · Open the model → try the original

1 / 8

Start with the trained model

Weights are learned numbers. They already encode capabilities and behaviours, including when the model tends to refuse.

Qwen3-1.7B

Example open-weight model used in this walkthrough.

In the notebook, choose Qwen3-1.7B, Granite 3.3-2B-Instruct or SmolLM2-1.7B-Instruct. The recorded examples here use Qwen3.

Weights

Learned numbers that determine the computation.

Tokenizer

Converts text into token IDs and back.

Chat template

Formats the conversation for this model.

Request → original model → baseline answer

Our generation setup

System message
You are a helpful assistant.
Decoding
Greedy · thinking disabled
Output cap
64 tokens · 32 in reduced CPU mode
Runtime
CUDA, Apple MPS or CPU

First check what the original model actually does. It may already answer some of the requests or make ordinary mistakes.

Input
An open-weight model checkpoint
Operation
Load weights and tokenizer
Output
The original model, ready to answer

Process state: Original weights · no edit

Sources and what this visualiser represents

Geometry is an illustration with invented coordinates. Prompt texts and answer panels come from the course materials and one recorded run of Qwen3-1.7B with Heretic 1.4.0 on mps, bfloat16, 2026-09-27. Output is capped at 64 tokens. Other runtimes and versions can differ.

This is a walkthrough, not a live connection to a notebook. Browser controls do not edit a real LLM or generate new answers.

The result of abliteration is altered behaviour, with no guarantee that the model answers every request, answers correctly or retains all capabilities.