weight-space-debugger
/weight-space-debugger
Sprawdza punkty kontrolne sieci neuronowych, bada powtarzalne błędy modelu i testuje autoryzowane, odwracalne edycje wag.
- Autor
- khengari77 ↗
- Repozytorium
- weight-space-debugger ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 08.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 8
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 1
Brak sygnałów. Repozytorium zmienione 08.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Sprawdza punkty kontrolne sieci neuronowych, bada powtarzalne błędy modelu i testuje autoryzowane, odwracalne edycje wag.
Oryginalny opis (EN)
Inspect neural-network checkpoints, investigate reproducible model failures, and test authorized reversible weight edits. Use for PyTorch, transformer, or RL checkpoint comparisons, numerical anomalies, causal circuit investigations, and targeted repairs; supports checkpoint-only access without inference.
Jak działa
- Route by access and requested work
- Shared constraints and evidence gates
- Tools and conditional references
- Provenance and contribution claims
- Deliverables and stopping
Podgląd SKILL.md
--- name: weight-space-debugger description: Inspect neural-network checkpoints, investigate reproducible model failures, and test authorized reversible weight edits. Use for PyTorch, transformer, or RL checkpoint comparisons, numerical anomalies, causal circuit investigations, and targeted repairs; supports checkpoint-only access without inference. license: MIT --- # Weight-Space Debugger Build discriminating evidence about neural-network behavior. Treat weight patterns, attention maps, probes and an LLM's explanations as hypotheses until tested. Inspired by [Sparks of Weight-Space Programming](https://rikiyat.github.io/blog/weight-space-programming/) by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. The concept is to reason directly about network weights and their computation, including deliberate edits rather than optimizer-driven training. This skill applies that idea to debugging; it does not reproduce the authors' experiments or promise general model repair. ## Route by access and requested work Determine the available access independently of intervention permission: | Access | Start here | Valid output | |---|---|---| | Checkpoint only, or inference disallowed | [Static forensics](references/09-static-forensics.md) | Integrity, structure and aligned deltas; untested behavioral hypotheses | | Executable model | [Executable investigation](references/08-executable-investigation.md) | Reproduction, behavioral measurements and authorized causal tests | | API only | Behavioral reproduction and config/input comparisons | Behavioral evidence; no weight-level claims | Choose the least invasive mode that satisfies the request: - **Forensics** (default): inspect weights, existing traces, or ordinary inference. No parameter edits, activation interventions or training steps. - **Causal debugging**: ephemeral activation or parameter interventions on an isolated run. No persistent checkpoint edits. - **Weight surgery**: when requested or already authorized, test minimal reversible parameter edits on a copy; accept only after held-out success and preservation checks. Training diagnostics are conditional on the task and experiment restrictions, not implied by access to training code. Follow [training isolation](references/07-training-debugging.md): use a disposable model and separate optimizer, preserve baseline parameters, buffers, gradients, module modes and RNG, and skip backward/optimizer steps when prohibited. Never train the investigation baseline. ## Shared constraints and evidence gates - Infer model identity/revision, tensor semantics, allowed data/inference, resource budget and preservation targets from available artifacts. Continue with useful inspection when inference or data is missing. - Use trusted model code and safe serialization. The audit supports `.safetensors`; do not auto-load pickle or enable remote model code. Do not export private weights or inputs without authorization. - For behavioral failures, freeze preprocessing, decoding/environment, precision, seeds and scoring. Rule out input, evaluation and inference defects before blaming weights. Static forensics needs no behavioral failure or datasets. - For causal claims, register a prediction, run the actual intervention, and include a meaningful control. Follow the [evidence method](references/01-evidence-and-method.md); distinguish necessity, sufficiency and specificity. - Before interpreting neurons, matrices or checkpoint deltas, check orientation, ties and parameter symmetries using [weight-space reasoning](references/02-weight-space.md). Unaligned coordinate differences do not establish changed behavior. - For repairs, use diagnostic/development/sealed-test splits and protected behaviors. Compare no-op, non-weight and norm/sparsity-matched controls. Follow [editing and validation](references/05-editing-and-validation.md). Static audits do not require these repair gates. - Preserve pristine checkpoints. `functional_call` can mutate aliased parameters/buffers when the forward updates state: use a disposable instance for stateful or training runs. Never silently replace or deploy a model. - Respect runtime and memory budgets. Retrieve small slices and aggregate statistics rather than dumping large weights or activation caches into context. A reproducible failure is required for a behavioral root-cause claim; a controlled intervention is required for scoped causal support; held-out success and preservation are required to call a persistent edit a fix. Report unsupported conclusions as **Unresolved**, with the next useful experiment or missing access. ## Tools and conditional references - Checkpoint inventory: `python scripts/checkpoint_audit.py FILE.safetensors --output audit.json`. - Aligned checkpoint deltas: `python scripts/compare_checkpoints.py BASE.safetensors OTHER.safetensors --output delta.json`. - In-process helpers: `scripts/probe.py` for sampled activation/gradient summaries; `scripts/causal.py` for single-tensor capture/patching; `scripts/patching.py` for reversible overrides and footprints; `scripts/metrics.py` for exploratory paired bootstrap intervals. - Transformer work: [tokenizer, decoding and circuit guidance](references/03-transformers.md). RL work: [observations, actions and control](references/04-rl-and-control.md). - Tool selection and primary sources: [toolbox](references/06-toolbox-and-sources.md). Read installed-version documentation before architecture adapters or model conversion; verify logit parity after conversion. - Loop investigation: [worked blueprint](examples/transformer-loop-investigation.md). Helper demonstration: `python examples/toy_case.py` (synthetic known fault). Paths above are relative to this skill's directory; resolve them there even when working in another project. Load only the references relevant to the selected task. ## Provenance and contribution claims Disclose sources of hypotheses, methods and code before presenting results. If a fix was found in prior work, say so explicitly; do not describe it as independently discovered. Separate reproduction, adaptation, engineering refinements and new findings. Compare claimed additions with the source: inherited evaluation controls are not original contributions. Use **novelty not established** when significance has not been demonstrated. Keep failed interventions and evidence limits visible; a successful reproduction does not establish autonomous discovery ability. ## Deliverables and stopping Use [the incident template](assets/incident-report.md) and [experiment registry](assets/experiment-registry.json) for substantive investigations. Mark unavailable behavioral, split and repair fields as not applicable for static forensics; record zero model forward/backward passes. Report artifacts/revisions, sampling coverage, observed facts, hypotheses, experiments, limitations, resource use and optional patch/rollback details. Use **Observed**, **Correlational**, **Causally supported (scoped)**, **Replicated**, **Unresolved** and **Falsified** accurately. Stop when the task is answered, evidence or preservation fails, or the budget is exhausted. A localized unrepaired failure or a useful static audit is a valid result. Validate helper changes with `python -m pytest -q` and `python scripts/package_skill.py`. These checks do not establish agent reasoning quality: use the separate [hidden-fault evaluation cards](examples/skill-evaluation.md) for that assessment.
Pliki w paczce (8)
- .gitignore 1 KB
- LICENSE 1 KB
- MANIFEST.json 10 KB
- README.md 8 KB
- SKILL.md 7 KB
- pytest.ini 1 KB
- requirements-dev.txt 1 KB
- requirements.txt 1 KB
Wersje
W bazie nie ma innych skilli o nazwie weight-space-debugger.
Migawka bazy z 13.11.2026 · wersja paczki 94efa56e7659 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100weight-space-programming
Analizuje i naprawia wagi sieci neuronowych przez rozumowanie i edycje tensorów, bez treningu, np. przy pętlach rozumowania. - 100weight-space-programming
Analizuje, diagnozuje i bezpośrednio edytuje wagi sieci neuronowych, z porównaniami i wycofaniem zmian, bez treningu. - 100weight-space-reading
Odczytuje wagi sieci neuronowych bez przejść w przód, by ustalić, czego się nauczyła, i porównuje checkpointy w przestrzeni wag. - 100weight-space-surgery
Diagnozuje błędy sieci neuronowych i naprawia je małymi edycjami wag bez gradientu, np. pętle powtórzeń w modelach językowych. - 95weight-space-programming
Diagnozuje błąd sieci neuronowej, czytając wagi i aktywacje, i naprawia go małymi edycjami wag bez ponownego treningu. - 100weight-surgery
Diagnozuje błąd wytrenowanej sieci przez odczyt wag i naprawia go małą edycją, a poprawkę sprawdza na danych testowych.