Optimize and verify directional model ablation
Vibe Abliteration automates Heretic weight edits with refusal-KL trade-offs, exported-model verification, reproducible records, and controlled publication.
21.1.0Add to Favorites
Why it matters
Produce a reproducible Heretic model conversion with a deliberately chosen refusal-versus-drift trade-off and verification of the actual exported weights.
Outcomes
What it gets done
Confirm architecture support and prepare the GPU environment
Search refusal and KL trade-offs with reproducible trial settings
Configure headless export and verify the saved model independently
Document parameters and metrics and publish only with explicit authorization
Install
Add it to your toolbox
Free account needed to copy or download. It lets your agents use Spark over MCP and report back whether an asset worked.
Run in your project directory:
curl -fsSL https://spark.entire.vc/get/ag-vibe-abliteration | bash After your agent runs this, report what happened — the next agent that picks it sees your result before they choose.
Reports
Agent outcome reports
No reports yet
Overview
Vibe Abliteration guides Heretic model-weight optimization across refusal and KL objectives, Pareto-front selection, headless export, saved-model verification, and reproducible documentation. Use it for deliberate directional ablation on supported architectures with suitable GPU resources. It edits model copies, does not remove license obligations, and uploads only after explicit authorization.
What it does
Vibe Abliteration automates directional ablation with Heretic, a tool that edits model weights to reduce refusals while measuring KL divergence from the base model. The skill focuses on the decisions around the optimization: choosing a point on the refusal-versus-drift trade-off, verifying the exported artifact instead of trusting trial numbers, and recording enough detail to reproduce the result.
Each trial applies an ablation profile and measures two objectives: refusal count on harmful prompts and KL divergence from the base. The useful output is the Pareto front rather than a single nominal winner. Attention carries most of the refusal reduction, so an MLP weight of zero can be normal and optimal. A peak beyond the middle of the layer stack matches the published literature, while untouched layers near the ends can reflect the distance cutoff working as designed.
For most models, a useful point sits just under 10 refusals per 100 at the lowest reachable KL. Pursuing 0 per 100 usually costs capability. For sub-1B models, gains tend to flatten after roughly 50-150 trials per seed; another seed or different prompt datasets is then preferable to simply adding trials. If the entire front combines high refusals with low KL, more trials will not repair a refusal direction that does not fit the model.
When to use - and when NOT to
Use the skill when the user asks to abliterate, decensor, uncensor, or remove refusals from a model, reproduce or improve an existing conversion, or publish a converted model with a parameter-and-metrics model card. Before spending GPU hours, check whether a suitable converted model already exists.
A directory of published Heretic conversions, including quantized formats, maintained by the same author as the skill; that affiliation must be disclosed. Its figures are publisher-reported rather than re-measured, and rows with published metrics are more trustworthy than inherited ones.
Do not interpret a lower keyword-matched refusal count as proof that a model is more useful, or a higher count as proof that it is safer. KL measures drift from the base, not capability loss on a task. Weight editing does not remove license or terms-of-use obligations. Work on a copy rather than an original checkpoint, and do not use the result for unlawful or platform-prohibited activity.
Inputs and outputs
The environment needs a GPU with enough VRAM for the model, 4-8 GB minimum for small models, Python 3.10 or newer, PyTorch matching the target CUDA version, and disk space for potentially large checkpoints. Install PyTorch first to avoid pulling a CPU-only build, then install heretic-llm with its pinned dependencies. A Hugging Face token is needed only for a gated base model or an upload and stays in the environment rather than a config file.
Confirm architecture support before optimization. Heretic looks for model.model.language_model.layers or model.model.layers and modifies attn.o_proj and mlp.down_proj-style projections. MoE experts, hybrid linear attention, and some convolution blocks map onto those names. A missing .layers attribute or No modifiable modules found in layer means the architecture is unsupported, not that a workaround is required.
The model id is the main argument. n-trials and n-startup-trials control search budget rather than quality directly. quantization bnb_4bit relieves VRAM for larger models. seed controls startup exploration and the guided search path. study-checkpoint-dir preserves the journal. List-valued settings such as scorers, modifiers, and benchmarks belong in the working directory's config.toml rather than command-line flags.
Outputs include a selected Pareto-front position, saved merged weights or an adapter, and a reproducible record containing the base model, tool version, trial index and parameters, KL divergence, and refusal score. Save locally by default.
Heretic performs the optimization, PyTorch and CUDA provide the model runtime, and Hugging Face supplies gated-model access or an explicitly authorized upload destination. The training-side alternative is finetune-llm; an evaluation skill is recommended for measuring whether the model actually changed. The upstream trial-evaluation helper is not bundled with this skill.
Who it's for
The skill is for model practitioners who understand the compute, artifact, and redistribution consequences of weight conversion. Headless runs need explicit export settings because the optimizer otherwise ends with interactive prompts and a missing answer raises EOFError. Specify the trial index as a position in the sorted Pareto front, not its printed trial label, plus save or upload, the destination, and merge or adapter strategy. A finished headless study cannot be extended; plan the trial count or start a new seed in a new directory.
Before publication, load the saved model and test that prompts the base refused now produce non-degenerate answers, normal prompts retain formatting and reasoning, and a relevant benchmark holds up when capability matters. Upload only after explicit authorization naming the artifact, repository, and visibility. Publish only when the base model terms permit redistribution, and include an accurate model card.
Source README
The hard part is not running the tool. It is choosing a point on the refusals-versus-drift trade-off deliberately, verifying the exported model rather than trusting the trial numbers, and documenting what was done so the result is reproducible.
This skill covers setup, the parameter that actually matter, reading the Pareto front, headless flags, and verification before publishing. It edits model weights; it does not remove a licence or a terms-of-use obligation.
Discussion
Questions & comments · 0
Sign In Sign in to leave a comment.