Replace the process-wide buffer (test hook).

Resolve `(subcommand, args)` into an executable plan. Raises: ValueError on unknown subcommand or invalid arg FileNotFoundError when

Map a vLLM ``CompletionOutput`` to an OpenAI ``finish_reason`` (#333). Trusts the engine's own ``finish_reason`` when it is one Soup emits, and

Resolve a dotted ``layer_path`` (e.g. ``model.layers.5``) to a submodule. Each segment is either an attribute name (``getattr``) or an integer in

THE answer to "which modules does this MLX run train?" (#392). Both the trainer and the ``adapter_config.json`` writer must use it. They used

Read the documented env knobs, applying validation. `env=None` reads from `os.environ`.

ROUGE-N F1 (corpus average over total pair count). Pairs where either side has fewer than `n` tokens contribute 0.0 to the average — short-se

Extract a can, run ``soup train`` against the embedded config, optionally run the embedded deploy targets. Args: can_path: path to th

LIVE lm-eval-harness invocation (#211). Resolves either an explicit ``tasks`` list (lm-eval task names) or the ``suite`` profile to its lm-ev

Run a small CMA-ES-style search over simplex merge weights. Operator supplies ``eval_fn(weights) -> score`` (higher is better); we softmax N-

Run the live compilation (v0.71.13 #225). Dispatches by ``plan.optimizer``: DSPy (bootstrap_fewshot / mipro / copro), GEPA, or TextGrad. Each

Execute one full iteration synchronously. Pure with respect to time.

Build an :class:`UnlearnReport` from optional pre-computed evidence. Missing evidence per-metric falls through to ``_neutral_metric`` (OK sco

Evenly-spaced sample of up to ``n`` indices across ``[0, total)``. Even spacing (rather than ``data[:n]``) catches issues confined to the tai

Write a JSON report with the sweep + recommended LR.

Containment-checked safetensors load + scan. Raises ``ValueError`` if the dir is outside cwd or a symlink; raises ``FileNotFoundError`` if no

Score 1 - fraction of training rows overlapping any benchmark row. ``benchmark_corpus`` may be a sequence of strings OR a sequence of Mapping

Return [0, 1] educational-value score. Combines (a) log-scale length and (b) type/token ratio as a proxy for vocabulary breadth. Lightweight

Return [0, 1] toxicity score from a keyword baseline. A real Llama-Guard variant lands in v0.47.1 via ``[data-pro]``.

Load model, tokenizer, apply LoRA, create trainer.

Apply the LongLoRA S² shift to half the attention heads. Pure math kernel used by :func:`apply_longlora_forward_override`. Given an attention

Split text into (prefix, suffix). Default: split by whitespace word-count fraction. When ``tokenizer`` is supplied (v0.71.6 #167 — an HF mode

Reduce a metric series into headline numbers.

Run the pipeline: chunk → judge → active-prune → ForgeRow rows. The ``judge`` callable is invoked once per chunk; failures are swallowed at D

Isolated registry + branches dir under cwd, returns a push helper.

Return a friendly install advisory for ``name`` if the package is missing, else None. Always returns ``None`` for `wandb` / `tensorboard` / `

v0.35.0 #61 — attempt ``build_fn(picked.name)``, falling back to the next-highest-scored candidate on load failure. ``build_fn`` is a callabl

Update Elo ratings after a match where A scored `score_a` ∈ [0,1]. score_a == 1.0 → A won. score_a == 0.0 → B won. score_a == 0.5 →

One-step exponential moving average update. `prev=None` initialises to `sample`. `alpha` is the weight on the new sample (0 < alpha <= 1). Sm

In-place EMA of the reference model from the policy (v0.71.11 #160). ``ref.param = (1-α)·ref.param + α·policy.param`` for every shared parame

Upload ``folder_path`` to ``repo_id`` on ``hub``. Same lazy-import policy as :func:`download_repo`. Token resolution is left to the caller (e

Validate an ``emotion`` tag for ``family`` (currently Orpheus only). Raises if the family does not support emotion conditioning or the tag is

Schema-time gate for ``fp8_attention=True``. Rejects: - non-bool ``fp8_attention`` (defence-in-depth). - ``fp8_attention=True`` without `

Validate GaLore configuration and return list of error messages.

Validate the codebook entry dimension (residual-stream width).

Validate the two-sided clipping delta. Must be a finite float in ``(0, 1]``. Bool rejected (matches v0.30.0 Candidate policy).

Raise ``ValueError`` if the LongLoRA prerequisites are not satisfied. LongLoRA requires: * ``task='sft'`` (preference / RL trainers have a

Validate + canonicalise a quant candidate for the live measure loop. Accepts the closed :data:`MEASURABLE_QUANT_CANDIDATES` allowlist plus the

Validate ``activation_offloading`` config. Args: target: None / 'cpu' / 'disk'. backend: transformers / unsloth / mlx. de

Validate that the Ollama URL is localhost-only (SSRF protection). Raises ValueError if the URL is not a local address.

Validate the per-round pair count. Range: ``[10, 1_000_000]``. Below 10 pairs the DPO gradient signal is too noisy to be meaningful; above 1M

Validate a prompt-program path (cwd-contained, ``.py`` only, no symlink). Returns the realpath of the validated path.

Validate a ``module.path:function_name`` rollout-func spec. ``None`` passes through (field unset). Mirrors the v0.42.0 ``validate_prompt_stra

Normalise + validate an ingest-source name. Mirrors v0.41.0 / v0.50.0 / v0.51.0 / v0.62.0 validator policy: bool-first, non-string -> TypeErr

Validate a student model id.

Validate ``threshold`` is a finite positive float.

Schema-time gate for ``unsloth_bnb_4bit=True``. Promotes "Unsloth Dynamic 4-bit" from "inferable from backend+quant" to a native flag. The fl

Validate the vLLM server URL (SSRF protection). Reuses the same validation as the server provider: - Scheme must be http or https - HTTP

Verify a detached ed25519 signature. Returns ``False`` on any verification failure (bad signature, wrong key, malformed hex). Raises ``TypeEr

Atomic write of an Annex XI or XII doc to ``output_path``. ``fmt`` is ``markdown`` (default) / ``md`` for the markdown body, or ``pdf`` (v0.7

Atomic write — cwd containment + symlink rejection at target.

Write rows as JSONL, atomically. Returns the realpath.

Write the recipe YAML under cwd; returns the realpath written.

Auto-deploy a GGUF file to Ollama after export.

Try to infer chat template from soup.yaml in cwd.

Move a TRL batch onto the model's device. `TrainingArguments` picks CUDA whenever it is available, so a test that deliberately pins the MODEL

``on_cuda=False`` exercises the panel without needing a GPU: the fit decision is skipped, but the calibrated budget and its display line are b

A real wrapper over a real tiny checkpoint, with `setup()` called.

Render the ``task: distill`` config: student = draft base, teacher = target. Every user-supplied string is embedded via ``json.dumps`` — a JSON s

Resolve fp16/bf16 kwargs for GRPOConfig (v0.53.3 #128). Priority: - Non-CUDA device (CPU / MPS / XPU) → no mixed precision (both

Strip ANSI colour codes + collapse whitespace from CLI output. Under CI ``FORCE_COLOR``, Typer/Rich injects ANSI codes that split flag names

Extract the final answer from model output. Supports: - #### <answer> format (GSM8K style) - \\boxed{<answer>} format (math style)

Best-effort prompt text from a calib row (text / prompt / messages).

Format bytes into human-readable string.

Get the format specification string for the given format.

Attach tags + artifacts to an entry row.

Construct an LLM provider instance with a ``generate(prompt)`` method.

Load a model and tokenizer (reuses diff.py pattern).

Create a mock TrainingArguments.

Create a minimal SoupConfig for display testing.

Create a mock TrainerState.

Parse ``name=path`` adapter specs; cwd-contain each path.

Parse adapter name=path pairs from CLI flag. Returns dict mapping adapter name → path string. Raises ValueError on invalid format.

Convert dataset rows to GRPO format. GRPO expects each row to have a 'prompt' field (list of messages or string). Input can be: - messa

Convert dataset rows to reward model format. RewardTrainer expects each row to have 'chosen' and 'rejected' text fields. Input can be:

Read JSONL file and return list of dicts.