spicy-llm

name	spicy-llm
description	Use the local spicy_llm Ollama/Heretic setup for decensoring, refusal-rate, and erotica/censorship evidence tests.

spicy_llm Local Model Skill

Use this skill whenever the user asks Hermes/Claude to use, test, benchmark, or inspect the local spicy_llm model setup.

What This Repo Is

/Users/jleechan/projects_other/spicy_llm is a local abliteration/decensoring research repo. It intentionally captures prompts and outputs that stock models may refuse. Harmful, sexual, or otherwise normally refused content in transcripts is expected research evidence, not a problem by itself.

Local Ollama Server

Check the server first:

curl -fsS http://127.0.0.1:11434/api/tags | python3 -m json.tool
ollama list
ollama ps

Current expected endpoint:

http://127.0.0.1:11434

If the endpoint is down, start Ollama:

ollama serve

Before using the Heretic model, ensure it is pulled:

ollama pull svjack/gpt-oss-20b-heretic

For browser chat, start Open WebUI from the repo root:

./install.sh

Open http://127.0.0.1:3100 and select svjack/gpt-oss-20b-heretic:latest in the model picker. Do not use Open WebUI as proof that Ollama is healthy unless the Ollama API check above also passes.

Model Selection

Use this decensored model unless the user names another one:

svjack/gpt-oss-20b-heretic:latest

Use this for stock baseline comparisons:

gpt-oss:20b

Do not silently replace the requested Heretic model with gemma2, gemma3, or another local model just because it is smaller or already loaded.

Running Prompts

CLI:

ollama run --verbose svjack/gpt-oss-20b-heretic:latest "$(cat path/to/prompt.txt)"

HTTP API:

curl -fsS http://127.0.0.1:11434/api/generate \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "svjack/gpt-oss-20b-heretic:latest",
    "prompt": "Prompt text here",
    "stream": false
  }' | python3 -m json.tool

For A/B tests, run the same prompt against both gpt-oss:20b and svjack/gpt-oss-20b-heretic:latest.

Evidence Capture

Create a dated results directory:

mkdir -p results/YYYY-MM-DD_test-name

Recommended artifact names:

prompt.txt
stock_transcript.txt
heretic_transcript.txt
REPORT.md

For user-facing copies, create cleaned companion files such as:

story_only.txt
human_readable_story.txt

Preserve raw transcripts when they support a research claim. Do not silently sanitize or delete raw output unless the user asks.

Review Priorities

Flag:

• secrets or private tokens
• private personal data
• accidental large model artifacts
• misleading claims not supported by evidence
• README instructions that do not work from a fresh clone
• .gitignore behavior that does not match the comments

Do not flag:

• harmful or sexual prompt/output artifacts merely because they test censorship
• raw refusal-removal evidence merely because the output is normally disallowed

Current Known Beads

• jleechan-u7i: fix results/ allowlist behavior
• jleechan-rz0: make the repo reproducible from committed harness files
• jleechan-ob1: define sanitized or gate-friendly transcript artifact path
• jleechan-rsu: document local Ollama/Hermes usage for this repo