Seed python / numpy / torch before the model and adapter are built. Returns the seed applied so a caller can report it.

Single entry point for trainers — returns a transformers quantization_config object (or ``None`` for ``quantization='none'``). Lazy-imports t

Pull the model's chosen MCQ option letter (A–J) from ``output``. Three "explicit commitment" forms are considered together — a LaTeX ``\\boxe

Isolated cwd + registry DB for loop tests; returns a state factory.

Parse a task-arithmetic expression into signed ``TaskTerm``s. Grammar (NO ``eval``): ``expr := term (('+'|'-') term)*`` where ``term := [sign

Load a ``LoopState`` from disk.

Append a verdict to the history log under an exclusive file lock. Raises ``TypeError`` / ``ValueError`` on shape errors so the CLI layer can

Live DataInf-style blame runner. Executes the v0.57.0 plan and returns top-K row influences. The ``probe_fn`` closure returns ``(row_grads, p

``seed`` / ``data_seed`` kwargs for any ``TrainingArguments`` subclass. ``data_seed`` stays ``None`` when unset: HF reads that as "follow ``seed`

A SHIP evidence payload (task improves, no regression).

A streamed model over a real (tiny) on-disk Llama checkpoint.

Apply Cut-CE / FP8 / kernel-auto-compose features to ``model``. Returns a dict ``{feature_name: applied}`` so the caller can log the decision

Attach :class:`SoupPluginCallback` when any enabled plugin implements a hook. Returns ``True`` when a callback was attached, ``False`` otherwise

Measure the instrument's resolution from repeated BASE-model runs. Each element of ``runs`` is one ``{axis: score}`` map from an independent

Build a :class:`CanarySet` from training rows. Algorithm: 1. Cluster rows by token-set Jaccard distance. 2. Pick top-N rows from each

Generate a fresh ed25519 private key and return its PKCS8 PEM string.

Get GPU memory info.

Parse a LLaMA-Factory YAML config and return a Soup config dict. Returns a dict suitable for config_to_yaml(). Includes a ``_warnings`` key w

Compute citation precision / recall / F1. Precision = |predicted ∩ expected| / |predicted|. Recall = |predicted ∩ expected| / |expected|.

Validate ``name`` against ``SUPPORTED_HUBS`` and return canonical form. Mirrors v0.41.0 ``validate_optimizer_name`` policy: rejects non-strin

A tiny llama-shaped checkpoint. Shapes are block-size friendly.

Shard a tiny Llama as NF4 and build the streamed model over it. ``quant_device`` is pinned to the training device on purpose: CPU and CUDA ag

Parse a JSON array from LLM output, handling markdown code blocks. Delegates to soup_cli.data.providers._utils.parse_json_array to avoid circ

Attach :class:`DynamicCurriculumCallback` when ``curriculum_dynamic=true``. Returns ``True`` when attached, ``False`` otherwise. The schema-level

Build a peft config spec from Soup's schema LoraConfig. Returns: Dict with keys: - ``peft_cls``: str — class name to import from

Filter ``pairs`` by LLM-judge confidence. The judge scores ``chosen`` and ``rejected`` independently. Each weighted score is normalised to ``

Append a completed edit to the governor's history.

Score refusal-rate regression vs base. Score = 1 - 0.5 * |Δharmful_refusal| - 0.5 * |Δbenign_refusal|. ``lang`` selects the heuristic patter

Build a task wrapper through the REAL `setup()` path, streaming. ``device`` defaults to the real accelerator, because `TrainingArguments` pic

Parse sweep parameter strings into a dict of {key: [values]}.

Make ``commands/data_doctor.py``'s ``engine.resolve_tokenizer`` return ``tok``.

ANSI-stripped, whitespace-collapsed CLI output, safe to substring-match. Collapsing whitespace as well as stripping escapes is deliberate: Rich a

A real (tiny) Llama checkpoint on disk — no network. ``hidden_size`` is 64 and NOT 32, and that is load-bearing for the NF4 tests. bitsandbyt

Minimal SoupConfig YAML that exercises the mitigation gate.

Apply a knowledge edit live (v0.71.9 #194 / #197 / #203; v0.71.16 #250). ROME / MEMIT / AlphaEdit run the rank-1 weight-edit kernel; ``grace``

Attach :class:`ReLoRACallback` when ``training.relora_steps`` is set. Returns ``True`` when a callback was attached, ``False`` otherwise. The

Map ``score in [0, 1]`` → ``OK`` / ``MINOR`` / ``MAJOR``. Rejects bool (subclass of int — matches project bool-as-int policy), non-finite, an

Run the live adapter conversion described by ``plan``. Live since v0.71.21 (#228) for the ``hf-to-mlx`` / ``mlx-to-hf`` directions; the two `

Run the named rollout backend and return its normalised result. Live since v0.71.21 (#125): - ``openenv`` resolves ``rollout_func`` (``modul

Log a single metrics row for the given run.

Build a ``judge(prompt) -> {'text': str}`` callable for ``provider``. Args: provider: one of ``"ollama"`` / ``"anthropic"`` / ``"vllm"``.

Exact concatenation-based task arithmetic over MIXED-rank LoRA adapters. For each LoRA module shared (as a complete A/B pair) across every adapte

Convert unified messages format back to a specific format. Input: {"messages": [{"role": ..., "content": ...}, ...]} Output: dict in target f

Compute stats and validate dataset.

Validate a steering strength multiplier. Bool-rejected (bool is a subclass of int), NaN/Inf-rejected via ``math.isfinite``, bounded ``|streng

Enable Rich progress bars for HuggingFace Hub file downloads.

A real (tiny) Llama checkpoint on disk: config.json + model.safetensors.

Validate a single generated example matches the expected format.

Wrap a trainer's ``trainer.train()`` call with activation offloading. Centralises the cwd-containment guard for ``activation_offloading="disk"``

Run pre-LoRA surgical patches (v0.39.0 Part D, multi-trainer in v0.40.6). Currently: Gemma4 ``ClippableLinear`` -> ``nn.Linear`` swap. Gated by

Build a validated :class:`AppleAdapterPlan` from raw CLI inputs.

Build the generation prompt for reasoning data. Args: count: Number of examples to generate. fmt: Output format (alpaca/sharegpt/

Estimate max batch size that fits in GPU memory. Conservative estimate — better to start smaller and gradient accumulate.

Insert ``num_new_blocks`` zero-init transformer layers at the end. v0.53.4 #83 — live implementation. Clones the last ``num_new_blocks`` deco

Greedy / temperature-sampled blend generation. Returns the full seq. Single-sequence (``B == 1``): the EOS stop inspects row 0, so a batched

Get full details of a single run. Supports prefix matching.

Load dataset for training. Returns dict with 'train' and optionally 'val' keys. Supports: - Local files (.jsonl, .json, .csv, .parquet, .txt)

Live canary verdict for a merged adapter (v0.71.4 #172). Returns the merged adapter's OK / MINOR / MAJOR verdict by comparing per-prompt cana

Render the shell script that takes the trained model to the target. The script is intentionally a stub: it prints the planned ``soup serve``

Return the HF token following the documented precedence. Order: 1. ``explicit`` argument (for tests / programmatic callers). 2. ``HF_TOKE

Create a fake adapter directory with adapter_config.json.

Return the MLP down-projection for decoder ``layer``. Llama: ``mlp.down_proj`` (nn.Linear, weight ``[out, in]``). GPT-2: ``mlp.c_proj`` (Conv

Build a minimal blame plan for testing.

A real ``transformers`` tokenizer — so ``apply_chat_template`` is the genuine Jinja renderer, not a mock that would agree with anything. Buil

Apply a vector bank to a running ``soup serve`` instance (v0.71.12 #221). Reconstructs the shared projection ``P`` from ``projection_seed`` and b

Generalized DPO loss kernel (v0.53.2 #135). Three variants: * ``standard`` — closed-form DPO: ``-log σ(β·(Δπ - Δref))``. Requires refe

``(bf16, fp16)`` for ``TrainingArguments`` on ``device``. bf16 where the card has it, fp16 where it does not, neither on CPU. Cannot regress

Build labels where only assistant tokens contribute to loss. Args: messages: Chat messages list (``{"role": ..., "content": ...}``).

Build the live echo-trap HF Trainer callback (v0.71.11 #240). Lifts the v0.70.0 ``NotImplementedError`` stub. Validates every input at the pu

Build the model structure on ``meta``: no weight storage is allocated. Under ``quant='nf4'`` the decoder linears are replaced with ``bnb.nn.L

Return a list of human-readable problem strings (empty = OK). Hard incompatibilities are emitted as plain strings; advisories use a leading `

OK / MINOR / MAJOR classification on a [0, 1] score (higher better).

Return normalised sampler weights per bucket. Buckets with no recorded samples fall back to the uniform prior. Buckets with higher mean loss

Generate Ollama Modelfile content. Args: gguf_path: Path to the GGUF model file. template: Template name (key in OLLAMA_TEMPLATES