Write a minimal loadable LoRA adapter dir; return its path string.

Seed a registry entry (and optional artifact / lineage) in a temp DB.

A PRMScorer with an injected fake model + tokenizer (no network). Fake model: hidden_states[-1] = arange(T) broadcast over H, and reward_head

Apply a (covariance-preconditioned) rank-1 update in place. Makes ``down(key) == old + delta`` exactly, so the down-proj now produces the opt

A real SFTTrainerWrapper over a real tiny checkpoint.

Inject the gpt-oss reasoning-effort control tag into the system message. Returns a NEW list (input is not mutated; mirrors v0.33.0 #47 immutabili

Attach a file at ``path`` to a registry entry as ``kind``. Returns the inserted artifact rowid, or None on lookup failure. Raises ``ValueErro

Build argv that wraps ``script_args`` with ``accelerate launch``. When ``num_processes == 1`` the launcher wrapper is skipped and ``script_ar

Build the hardware-info-only telemetry payload. The payload contains NO user data, dataset paths, model names, or config contents. Documented

Suggest ``target_modules`` values for the chosen ``base`` model. When ``base`` names a model whose HF config is in the local cache, v0.71.1 #

Build a ``FailureReport`` with the overall verdict computed for you.

Rank each canary's loss against the control distribution. ``percentile`` = fraction of controls STRICTLY cheaper than the canary. Low percent

Compute per-feature mean activation diff and return top-K changes. Both arrays must have shape ``[N_tokens, N_features]``. ``top_k`` is clamp

Return a list of ``{kind, snippet}`` PII hits. Scans only the first ``_PII_SCAN_CAP`` chars of ``text`` to keep regex finditer cost bounded r

Estimate per-run cost in USD; None when the GPU is not priced. Returns None for CPU / MPS / unknown devices so callers can render a "—" inste

Export a HuggingFace model as a UD / IQ / Apple-ARM GGUF. Three-stage pipeline: 1. ``convert_hf_to_gguf.py`` → ``f16.gguf`` 2. If ``flavo

Extract every citation id from ``text`` for the given ``style``. Returns a tuple of IDs in encounter order. Duplicates are preserved so the c

Mark run as completed and fill summary fields. Also computes an informational per-run cost estimate based on the device_name captured

Apply per-family emotion templating to a chat-message list. Returns a NEW deep-copied list of message dicts (caller's list and its nested val

Extract per-sample token lengths from a tokenized dataset. Falls back to a ``"length"`` column when ``input_ids`` is absent. A sample with ne

Load per-prompt response rows from a JSONL file under cwd. Each row should carry at least ``{item_id, correct}``. Malformed rows are silently

Parse an OpenAPI 3.x ``dict``. Returns (endpoints, warnings).

Bucket ``value`` by its percentile rank within a rolling ``window``. v0.71.5 #149 — replaces step-mod round-robin with a difficulty-signal bu

Pick mixed-precision mode for a model + GPU. - cc < 6.0 → ``"no"`` (Pascal lacks reliable fp16 tensor cores) - cc < 8.0 → ``"fp16"`` (no bf

Build a `DelinearizePlan`. Raises on bad inputs.

Run one short proxy training and return the observed eval loss. Args: weights: Per-dataset interleave probabilities (simplex). da

Return the PEM-encoded public key for an ed25519 private key.

Append one entry. Never raises on serialisation issues.

Render a plain-text time-series of bucket weights over training. Each row is one recompute step; each column is one bucket. Output uses ASCII

Render a complete `soup.yaml` from a 5-answer dict. Required keys: base, dataset, task, quantization, epochs. Optional: output (default `./ou

Render the pre-push hook script body — no I/O, deterministic.

Apply a calibrated defection probe for ``base`` to ``activations``. ``activations`` must be a 2D array ``[N_tokens, hidden_dim]``. When ``we

Save an evaluation result.

Score a single task output against the expected answer.

Per-trajectory repetition score. Returns the fraction of n-grams whose count exceeds 1 (the "repeating n-grams" rate). Range ``[0, 1]``. 0 =

Normalise + validate a citation-style name. Mirrors v0.41.0 / v0.51.0 / v0.61.0 validator policy.

Validate and normalise a reward-hacking detector name. Returns the canonical (lower-cased) name on success. Raises ``ValueError`` with an act

Validate Ollama model name. Returns (is_valid, error_message).

SSRF-hardened OTLP endpoint validation. Rules (mirrors utils/hf.py): - scheme must be http/https - plain http only permitted for loopback

Schema-time gate for ``task='prm'``. Rejects: - non-PRM task (the function is intended to be called only when ``task == 'prm'``; defenc

Normalise + validate a steering-method name. Mirrors v0.41.0 / v0.51.0 / v0.61.0 validator policy: bool-rejected, null-byte-rejected, oversiz

Normalise + validate an unlearn-method name. Returns the canonical (lowercase) form. Mirrors v0.41.0 ``validate_optimizer_name`` / v0.51.0 ``

Decide whether ``repo_id`` can be loaded. Decision matrix: - Unknown repo: trust-on-first-use. Record and return ok=True. - Known repo,

Return a copy with ``status`` set + ``updated_at`` refreshed.

Generate examples using a local OpenAI-compatible server (soup serve, Ollama, etc.). Unlike the 'openai' provider, no API key is required. Connec

Walk user through config creation.

Helper: return a KTOTrainerWrapper with trainer pre-injected.

Resolve ``(key, endpoint)`` from explicit args + ``SOUP_POSTHOG_*`` env. v0.53.10 #154 — adds env-var overrides for the bundled defaults so users

v0.72.0 BETA — layer streaming. The resident base load NEVER happens. Builds the skeleton on ``meta`` (``accelerate.init_empty_weights``),

Build a closure that returns (row_grads, probe_grad).

Attach an artifact to an entry. Returns the artifact row id. ``enforce_cwd`` (on by default) rejects files outside the cwd captured a

Record a ``child → parent`` lineage edge. Rejects self-references and indirect cycles (e.g. A→B→A) so BFS walks in :meth:`get_ancesto

Convert attention-projection linears to FP8 training modules. Live since v0.71.21 (#141). Walks ``model.named_modules()``, collects every ``n

Add configured tokens to ``tokenizer`` and resize ``model`` embeddings. The helper is intentionally shared by every text trainer so ``data.ad

Build a live per-token gating kernel for MoLE dispatch (v0.71.12 #222). Returns a ``torch.nn.Module`` whose forward maps a ``[..., hidden_dim]``

Meta skeleton -> extras -> LoRA -> streaming. No resident base load.

Refuse pickle / PyTorch-classic weights when ``strict=True``. Args: model_dir: cwd-contained model / adapter directory. strict: w

RAM is Tier 1; disk is overflow only; spinning rust is refused. plan 2.3: streaming from NVMe measured 2-11 TFLOPS against multiples of that

OK / MINOR / MAJOR by absolute score (5% / 20% bands). Boundary semantics: exact 0.05 → MINOR; exact 0.20 → MAJOR.

Compute a pre/post diff report from explicit responses + oracle labels. All three lists must be the same length and contain str. The oracle entry

Compute coherence scores for a list of texts. Coherence is measured by: - Sentence-level structure (has proper sentences) - Word repetiti

Stream each FSDP shard and write a single consolidated ``.safetensors``. Loads one shard at a time via ``torch.load(weights_only=True)`` (no

Decide whether ``metric`` regressed past the configured tolerance. Uses the paired-bootstrap 95 % CI of the delta. Higher-is-better metrics r

Return the longest prefix shared by >= min_frequency of rows. Empty / single-row / no-overlap fall through to "" except the trivial single-ro

Check if Ollama is installed, return version string or None.

Per-license risk check for a target deployment. Decision matrix: - Unknown license -> ``warn`` (operator should review). - Non-commercial

Get all metric rows for a run, ordered by step.

Recommend the next method given accumulated state. Switching rules: 1. ``alphaedit`` is the survival-mode method — never switched away.

Pair up/down responses to the same prompt into DPO training rows. For each prompt that has at least one ``up`` AND at least one ``down`` thum

Read back the manifest from a bundle tarball. Containment-checked; symlink at the bundle path is rejected via the shared helper. Raises ``Fil

Best-effort check whether ``name`` matches a known VLM family. Returns ``False`` (never raises) on any of: non-string, empty, null byte, leng

Deterministic k-means++ over ``vectors``; returns integer labels.

Load an audio file as mono float32 at ``target_sr``. Lazy-imports soundfile (friendly ImportError naming the install). Off-rate audio is resa

Load model + tokenizer via unsloth FastLanguageModel with LoRA already applied. Returns (model, tokenizer) — model already has LoRA adapters atta

Load an ``AutoModelForCausalLM`` + tokenizer, optionally with a LoRA adapter. Returns ``(model, tokenizer, device)``. ``model`` is ``.eval()``-ed

Load an operator-supplied calibrated probe ``(W, threshold)`` from disk. Replaces the synthetic seed-derived placeholder with a real weight file

Load the dataset registry. Returns empty dict if file doesn't exist.

Return (canonical_label, usd_per_hour) for a device name, or None. Matches case-insensitively against `_GPU_RATE_TABLE`. Returns None when th

Parse an Unsloth .ipynb notebook and return a Soup config dict. Extracts parameters from function calls using AST parsing only. Returns a dic

Select the high-information subset of items per profile.

Append an event. Returns the event's monotonic cursor id.