crack-open-models-dgx-spark — Claude Code skill: what it does
The crack-open-models-dgx-spark skill for Claude Code designs and implements model behavior interventions on a DGX Spark cluster. Use it to modify refusal mechanisms in open-weight models.
MIT ★ 73 updated 2026-10-22
npx skills add jiayuqi7813/crack-open-models-dgx-spark --skill crack-open-models-dgx-spark -a claude-codeWhat crack-open-models-dgx-spark does
This skill automates architecture analysis, intervention evaluation, and deployment for open-weight models. It uses python3 scripts to perform checkpoint inventory and pipeline planning tasks.
A two-node DGX Spark cluster is required for operation. All heavy inference, checkpoint loading, and evaluation must run on the DGX pair, while the local machine handles code, manifests, and small metadata results.
What it helps with
- analyze model architecture constraints
- mitigate model refusal behavior
- validate intervention integrity after quantization
Example prompts for the agent
- python3 skills/crack-open-models-dgx-spark/scripts/new_run.py --model-id ORG/MODEL --objective refusal-behavior-edit --nodes dgx-spark-1,dgx-spark-2
- python3 skills/crack-open-models-dgx-spark/scripts/checkpoint_inventory.py /path/to/checkpoint --output runs/MODEL-SLUG/checkpoint-inventory.json
How to install crack-open-models-dgx-spark
1Skills CLI (Vercel)
npx skills add jiayuqi7813/crack-open-models-dgx-spark --skill crack-open-models-dgx-spark -a claude-code2Manually: put the files in place
~/.claude/skills/crack-open-models-dgx-spark/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/crack-open-models-dgx-spark/
1Skills CLI (Vercel)
npx skills add jiayuqi7813/crack-open-models-dgx-spark --skill crack-open-models-dgx-spark -a codex2Manually: put the files in place
.agents/skills/crack-open-models-dgx-spark/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add jiayuqi7813/crack-open-models-dgx-spark --skill crack-open-models-dgx-spark -a cursor2Manually: put the files in place
.cursor/skills/crack-open-models-dgx-spark/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install jiayuqi7813/crack-open-models-dgx-spark crack-open-models-dgx-spark2Manually: put the files in place
.github/skills/crack-open-models-dgx-spark/SKILL.md1Skills CLI (Vercel)
npx skills add jiayuqi7813/crack-open-models-dgx-spark --skill crack-open-models-dgx-spark -a gemini-cli2Platform CLI
gemini skills install https://github.com/jiayuqi7813/crack-open-models-dgx-sparkManually: put the files in place
.devin/skills/crack-open-models-dgx-spark/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about crack-open-models-dgx-spark
How to install crack-open-models-dgx-spark in Claude Code?
Run npx skills add jiayuqi7813/crack-open-models-dgx-spark --skill crack-open-models-dgx-spark -a claude-code. Reference the ~/.claude/skills/crack-open-models-dgx-spark/SKILL.md file for details.
What does crack-open-models-dgx-spark do?
It manages the lifecycle of model behavior interventions, including architectural forensics and verifying intervention performance on dedicated hardware.
What is required to run crack-open-models-dgx-spark?
A two-node DGX Spark cluster is mandatory, along with python3 to execute the provided inventory and planning scripts.
Source text
SKILL.md · MIT
--- name: crack-open-models-dgx-spark description: Analyze an open-weight model and design, implement, evaluate, deploy, or publish the best architecture-specific refusal-behavior intervention on the local two-node DGX Spark cluster. Use for CRACK or abliteration work, derivative-method reconstruction, model-architecture and quantization forensics, runtime sidecars, native checkpoint edits, training decisions, matched Base/candidate evaluation, or reusable dual-DGX delivery. Do not use as a fixed recipe or when the request is only ordinary model serving with no behavioral modification. --- # Architecture-first model intervention on DGX Spark Build a model-specific intervention from evidence. Reuse the research system, not another model's direction, tensor names, layer range, codec, hook, or strength. ## Operating boundary 1. Read the workspace `AGENTS.md`, then the cluster README and validation record it names. Probe both nodes before changing services. 2. Keep checkpoint download, loading, editing, capture, inference, and large evaluation on the DGX pair. The Mac may hold code, manifests, small metadata, exported activations, and text results only. 3. Preserve the pinned source. Every candidate must be reversible and separately identified. A runtime intervention is not an edited checkpoint. 4. Pin immutable model, tokenizer/template, runtime, dataset, judge, and code identities before comparing results. 5. Do not make the user choose an implementation route when architecture evidence can decide it. Investigate, select the best route, and report the reason. Create the run ledger at the start, even if the runtime route is still unknown: ```bash python3 skills/crack-open-models-dgx-spark/scripts/new_run.py \ --output runs/MODEL-SLUG/intervention-run.json \ --model-id ORG/MODEL --revision COMMIT \ --objective refusal-behavior-edit \ --backend auto --nodes dgx-spark-1,dgx-spark-2 ``` The ledger is the resumable source of truth. Record evidence paths and hashes, not narrative progress alone. ## Core workflow ### 1. Discover the actual system Read [references/architecture-discovery.md](references/architecture-discovery.md). Inspect current primary sources and the exact checkpoint metadata before downloading optional derivatives or writing hooks. On a DGX-local checkpoint, run: ```bash python3 skills/crack-open-models-dgx-spark/scripts/checkpoint_inventory.py \ /path/to/checkpoint --output runs/MODEL-SLUG/checkpoint-inventory.json python3 skills/crack-open-models-dgx-spark/scripts/plan_pipeline.py \ runs/MODEL-SLUG/checkpoint-inventory.json \ --output runs/MODEL-SLUG/pipeline-plan.json ``` Treat the generated plan as a hypothesis. Confirm its writer paths, collectives, stateful branches, quantization, special heads, and runtime support against source code and a real minimal forward pass. Update the ledger's architecture fingerprint and compatibility decision. Do not use model branding as architecture identity. Two similarly named releases may differ in layer topology, hidden size, quantization, custom code, speculative heads, or compatible runtime. ### 2. Establish a clean Base Before behavior modification, prove that the exact Base loads with the intended TP=2 runtime and passes the smallest real request. Freeze: - checkpoint and shard identity; - tokenizer, chat template, generation defaults, and served model name; - runtime image/source, quantization kernels, TP collective order, KV format, context, thinking/parser flags, and speculative path; - service health, output determinism, memory headroom, and known unsupported features. Do not debug an intervention on top of an unverified Base. ### 3. Acquire the behavioral signal Read [references/method-routing.md](references/method-routing.md) and choose the strongest available evidence source: - **Reference differencing:** when a credible derivative and exact high-precision Base exist, fetch indexes, headers, and the smallest exact tensor ranges first. Recover and validate the dominant change without downloading the whole derivative unless necessary. - **Paired activation discovery:** when no trustworthy reference exists, capture semantically paired target/control prompts at architecture-defined writer and residual surfaces, with train/selection/untouched confirmation splits. - **Hybrid:** use derivative differencing for the initial direction and paired activations to validate sign, localization, transfer, or residual correction. - **Training:** use only when measured evidence shows that closed-form or reversible runtime methods cannot meet the requested behavior/capability tradeoff. The direction is model- and surface-specific. Validate rank structure, sign, holdout separation, cross-layer consistency, branch coverage, and reproducibility. A high classifier score alone is not a deployable intervention. ### 4. Route the implementation by representability Select the implementation after measuring both mathematical effect and checkpoint-format expressibility: - **Native checkpoint edit:** use when target writer tensors and their codec are understood, zero-edit roundtrip passes, the intended edit survives re-encoding, and non-target invariants can be proved. - **Runtime sidecar:** prefer when low-bit packing swallows the edit, routed experts cannot be safely rewritten, the architecture exposes rank-summed writer outputs, or a reversible implementation is more faithful than weight surgery. - **Hybrid runtime correction:** use when projecting all relevant writer branches is insufficient and a measured residual boundary requires a bounded correction. Scope it by phase, token position, layer, and decode behavior. - **Adapter or training:** justify with concrete failure analysis, a reversible plan, and matched capability gates. Do not train merely because a previous model used training or because one linear candidate was weak. Zero-edit roundtrip is necessary but not sufficient. Always run an **edit-expressibility test** that decodes, applies the intended delta, re-encodes, decodes again, and measures how much of the target direction remains. ### 5. Build the smallest falsifiable candidate ladder Change one causal factor at a time. Candidate axes may include signal source, target branch, layer region, rank, projection versus reflection, strength, residual correction, quantization policy, or special-head handling. Select axes from measured architecture evidence, not a universal sequence. Each candidate must bind: - source fingerprint and runtime identity; - direction/correction artifacts and hashes; - exact targets, layers, branches, phases, and strengths; - implementation code/image hashes; - integrity results and evaluation plan; - whether the checkpoint changed. Discard candidates that fail integrity or cannot be faithfully represented. Do not score or promote partial files. ### 6. Evaluate with profiles matched to the request Read [references/evaluation-system.md](references/evaluation-system.md). Use the same Base/candidate checkpoint, runtime, prompts, decoding, and judge identity except for the intervention switch. Choose the narrowest honest profile: - **smoke:** integrity, startup, API, parser, deterministic probe, and a small frozen behavior set; - **behavior-only:** full requested refusal/compliance benchmark and benign controls; make no general-capability claim; - **preservation:** behavior plus coding, knowledge, math, writing, instruction/JSON/tool use, supported modes, and architecture-relevant long context; - **release:** preservation plus performance, stability, deployment, reproducibility, and publication checks. Keep candidate selection data separate from final confirmation. Report item-level transitions, errors, invalid labels, truncations, and judge disagreements. Do not convert `1 - ASR` into a literal lexical-refusal rate. ### 7. Freeze, deploy, and hand off Read [references/validation-release.md](references/validation-release.md). Start and stop the two-node stack in the project-required order, verify actual listeners and containers, and test the endpoints the user will use. Deliver the correct object: - edited checkpoint plus manifest; - unchanged checkpoint plus frozen sidecar and engine image; - adapter plus exact base identity; - or an evaluation report when no candidate passes. Publish weights from the DGX head, never through the Mac. A sidecar release should publish code, direction/correction assets, manifest, and reproduction instructions without relabeling unchanged Base weights as a new checkpoint. Update the cluster handoff documents with current live state, exact restore commands, identities, results, limitations, and the evaluation profile actually completed. ## Audit gates Use the audit helper as the ledger matures: ```bash python3 skills/crack-open-models-dgx-spark/scripts/audit_run.py RUN.json --mode plan python3 skills/crack-open-models-dgx-spark/scripts/audit_run.py RUN.json --mode candidate python3 skills/crack-open-models-dgx-spark/scripts/audit_run.py RUN.json --mode handoff python3 skills/crack-open-models-dgx-spark/scripts/audit_run.py RUN.json --mode publish ``` The audit checks evidence completeness, not scientific truth. Inspect the actual reports and outputs. ## Claim discipline Separate measured results, engineering inference, and untested areas. Say “no material regression on the tested suite” only after a preservation profile supports it. If only behavior testing ran, state that coding, writing, knowledge, tools, multimodal behavior, and long context remain unverified.
© jiayuqi7813 · MIT · jiayuqi7813/crack-open-models-dgx-spark