mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-11-17 04:11:35 +00:00
* chore(skills): remove red-team skills (godmode, obliteratus) from bundled catalog Anthropic's output classifier on claude-fable-5 (and likely other Claude models served through it) intermittently returns empty content for sessions whose system prompt advertises these skills. The bundled skills-catalog block is injected into every session's system prompt, so the descriptions - red-teaming/godmode 'Jailbreak LLMs: Parseltongue, GODMODE, ULTRAPLINIAN' - mlops/inference/obliteratus 'OBLITERATUS: abliterate LLM refusals (diff-in-means)' trip the classifier on EVERY session regardless of which skill is actually loaded, killing unrelated legitimate work (PR review, codebase audits, etc.). Measured impact (controlled, interleaved A/B, claude-fable-5 via OpenRouter, prompts differing only by the ~204 chars of these catalog lines, N=20 each): catalog lines present -> 19/20 (95%) blocked catalog lines absent -> 5/20 (25%) blocked Removing them ~quartered the block rate. Rewording the descriptions was not enough; the skills must leave the bundled catalog. - Delete skills/red-teaming/godmode and skills/mlops/inference/obliteratus - Drop their generated doc pages + catalog/sidebar entries (EN + zh-Hans) - Drop the godmode hand-written-page exception in generate-skill-docs.py * chore(skills): relocate godmode + obliteratus to optional-skills Rather than deleting outright, move both into optional-skills/ so they remain installable via `hermes skills install` while leaving the always-injected bundled catalog (which is what tripped Anthropic's classifier). - optional-skills/security/godmode (was skills/red-teaming/godmode) - optional-skills/mlops/obliteratus (was skills/mlops/inference/obliteratus) - regenerate optional-skills catalog + sidebar entries
33 lines
1.2 KiB
YAML
33 lines
1.2 KiB
YAML
# OBLITERATUS Abliteration Config
|
|
# Usage: obliteratus run this-file.yaml
|
|
#
|
|
# This is for reproducible, version-controlled abliteration runs.
|
|
# For one-off usage, the CLI flags are simpler.
|
|
|
|
# Model to abliterate
|
|
model:
|
|
name: "meta-llama/Llama-3.1-8B-Instruct"
|
|
dtype: "bfloat16" # float16, bfloat16, float32
|
|
quantization: null # null, "4bit", "8bit"
|
|
device: "auto" # auto, cuda, cuda:0, cpu
|
|
|
|
# Abliteration method and parameters
|
|
abliteration:
|
|
method: "informed" # See SKILL.md Step 4 for all 13 methods
|
|
n_directions: null # null = auto-detect, or integer (e.g., 8)
|
|
regularization: 0.0 # 0.0-1.0, fraction of original to preserve
|
|
refinement_passes: 1 # Iterative passes (increase for self-repair)
|
|
norm_preserve: true # Keep weight norms intact after projection
|
|
|
|
# Output
|
|
output:
|
|
directory: "./abliterated-models"
|
|
save_metadata: true # Save abliteration_metadata.json alongside model
|
|
contribute: false # Save community contribution data
|
|
|
|
# Verification
|
|
verify:
|
|
enabled: true
|
|
test_prompts: null # null = use built-in test prompts
|
|
compute_perplexity: true
|
|
compute_kl: true
|