SHIP iff ``ppl_final / ppl_original - 1 <= tolerance``. ``decide_ship`` (soup ship) would trivially reject every shrink because pruning alway

Detect whether a model uses Mixture of Experts architecture. Checks model config for known MoE indicators.

plan P5: logits, not weights, OOM you first on a small card. Measured (GATE 2), not derived: the loss path holds ``LOGITS_BYTES_PER_ELEMENT``

First-Fit-Decreasing bin packing. Args: lengths: per-sample sequence lengths. max_len: maximum sum of lengths in any bin. Re

Per-row series of a single named metric for a run (v0.55.0). Used by ``soup eval against`` for run-vs-run paired-bootstrap CI. Return

Create the loop.yaml state file. Refuses to clobber unless ``force``.

Read the most recent ``limit`` entries, newest first. Returns an empty list when the file does not exist. Malformed lines are skipped silentl

Load a reward function by name or from a custom Python file. Args: reward_fn_spec: Either a built-in name ('accuracy', 'format', 'verifia

Weighted average per layer over the intersection of names.

Build a leave-one-out ablation plan (v0.57.0). The plan splits the dataset into ``num_shards`` equal shards. ``feasible`` is False when the b

Build a `ConsolidationPlan`. Raises on missing shards or bad output.

Return the HF endpoint, validating ``HF_ENDPOINT`` when set. Strips one trailing slash, rejects non-HTTPS remotes (localhost HTTP is allowed

Run the BO loop over the validated plan. Args: plan: A :class:`MixOptimizationPlan` from :func:`build_optimization_plan`.

Validate an operator-supplied steering-vector name. Returns the value unchanged on success. Closed regex allowlist (alphanumeric + ``._-``, l

Write an importable rollout module; return the 'module:fn' spec.

Run post-LoRA surgical patches (v0.39.0 Part D, multi-trainer in v0.40.6). Currently: 3-D fused-MoE expert dropout strip. Architecture-detected v

Assemble the bundle. Returns the in-memory manifest. Raises ``ValueError`` on: - output / model_dir / dataset_dir outside cwd (containment)

Factory: return the ``format_row`` function appropriate for ``data_cfg``. v0.53.2 #137: when ``training_cfg.reasoning_effort`` is set the gpt-oss

Build a full N×N interference matrix from per-pair losses. ``losses`` keys are ``(target_adapter, co_loaded_adapter)`` tuples and values are

Return β at ``step`` for the chosen schedule. Endpoint contract (matches HF lr_scheduler convention): step ≤ 0 → ``beta_start``

Pick a quantization tier from model size vs available VRAM. v0.53.1 #82 — when ``prequantized`` is set (e.g. ``"gptq"`` / ``"awq"`` / ``"hqq:

Detect available device. Returns (device_string, human_name).

Return a list of findings, ordered by severity (critical first).

Greedy cosine near-duplicate removal over L2-normalized ``vectors``. Rows are visited in order. The first row of any near-duplicate cluster i

Return entries filtered by name/tag/base/task.

Load a JSONL dataset for advise, with cwd containment + symlink reject. Mirrors v0.53.7 #106 TOCTOU policy: ``os.lstat`` on the raw path BEFORE

Parse an Axolotl YAML config and return a Soup config dict. Returns a dict suitable for config_to_yaml(). Includes a ``_warnings`` key with a

Interleave a seeded sample of ``replay_rows`` into ``new_rows``. Interleaved, NOT appended. ``new + replay`` would put every replay row in on

Paired-bootstrap (lower, upper, mean) of ``candidate - baseline``. Standard paired-sample bootstrap with replacement at the row level (preser

Parse a sweep YAML payload.

Pick a known-good draft model for the target. A draft the user trained themselves via ``soup draft distill`` (recorded in the local registry,

Prepare a model for Quantization-Aware Training. Inserts fake quantization operators into the model so that forward passes simulate the effec

Render a ready-to-train ``soup.yaml`` for ``profile`` + ``base``. Returns the YAML *text*; callers handle write-out with their own containmen

Resolve the HF Trainer `report_to` value from CLI flags + --tracker. Mutual-exclusion: only one of (wandb, tensorboard, tracker) may be set.

Convert an OpenAI chat-completions payload to Anthropic Messages shape. - The (single) ``system`` message becomes the top-level ``system`` field.

Validate a battery name. Case-insensitive; returns canonical form.

SSRF protection for judge API base URL.

Schema-time gate for ``vision_grpo=True``. Rejects on: - task not in {'grpo', 'ppo'} (vision RL is only meaningful for RL); - modality !=

SSRF-hardened webhook URL validator (returns the canonical URL). ``allow_private_hosts`` defaults to ``False`` — the safe webhook policy that

Build a test FastAPI app via _create_app (lazy fastapi import).

Write a minimal LoRA-shaped adapter (safetensors + config).

A tiny real-torch model with q/k/v/o projections + an MLP linear.

Synthesize a 2-layer Llama-shaped model dir (numpy safetensors). No torch / no download — exercises the streaming scanner end-to-end.

Energy-Based Fine-Tuning loss kernel (v0.53.2 #135). EBFT treats per-token logits as energies (lower = more probable) and penalises high-ener

Stash the multipack sampler config on ``trainer``.

Build the bundle dict. Always returns; never raises.

Validate inputs and return an immutable plan.

Build the live reward-hacking HF Trainer callback (v0.71.11 #235). Lifts the v0.70.0 ``NotImplementedError`` stub. Validates every input at t

Build the live mid-epoch RL checkpoint callback (v0.71.11 #238). Lifts the v0.70.0 ``NotImplementedError`` stub. ``output_dir`` is the traine

Fit a steering vector from contrastive pairs and persist it. Validates method + name FIRST (so a typo surfaces before the model load), then r

The #1 'model never stops generating' bug. MAJOR when >=50% of rows have at least one assistant turn whose trained span never trains an EOS/EO

Decide whether a list of licenses can be combined in one merged adapter. Decision rule: - Normalise + dedup the input. - If any license i

Classify the task using keyword + structural signals. Pure-Python, no ML. Returns one of ``TASK_CATEGORIES``. The goal string (when supplied

Return the alpha-quantile of the calibration scores. With ``alpha=0.1`` and a well-calibrated set of judge confidence scores in ``[0, 1]``, p

Best-effort license auto-detection for an adapter / model directory (#187). Resolution order: ``adapter_config.json`` -> ``config.json`` -> ``REA

Walk ``template``'s Jinja AST and return the set of message fields. Detects both attribute-style ``{{ m.role }}`` and subscript-style ``{{ m[

Patch ``torch.cuda`` in place; the resolver imports torch lazily.

K unique canaries, deterministic in ``seed``.

Return the bundle metadata for `name`, or raise ValueError.

Return True when ``path`` resolves inside ``base``.

Load a YAML/JSON spec from disk with cwd containment + size cap. Symlinks are rejected (TOCTOU defence, mirrors v0.45.0 Part E policy).

Build a real CUDA ``probe_fn`` for :func:`pick_batch_size`. Returns a closure that, given a candidate batch size ``B``, runs ONE forward + ba

Prepare a distillation dataset from prompt-heavy traces (v0.71.13 #226). For each prompt in ``plan.traces_path`` the teacher model is called once

Parse a PEM string/bytes into an ed25519 private key. Rejects non-ed25519 keys (RSA / EC / etc.) with a clear ``ValueError`` so operators can

Record ``target -> draft_dir`` so ``serve --auto-spec`` can find it. The target key is lower-cased to match ``spec_pairing.pick_draft_model``'s

Render the workflow YAML body. Deterministic, no I/O. Every path is validated to stay under the repo root and shell-quoted before it reaches

Run all tests in ``spec`` against operator-supplied responses. ``evidence`` maps each test ``name`` to the list of responses (one per prompt)

Harvest the latest DPO pairs and train them (v0.71.13 #229). The flow: read ``last_train_at`` from the SQLite ``state`` table; if a prior tra

Search recipes by keyword, task, or model size.

Return a point-in-time copy of all metrics.

Return {zero_shot, few_shot, rag} baseline scores in ``[-1.0, 1.0]``. When ``model`` is supplied (and torch + the model load succeed) this runs

Tokenise a composed RAFT example into a pre-tokenised training row. Returns ``{input_ids, attention_mask, labels, loss_weights}`` (python lis

Validate dataset paths: containment + dedup + bounds. Args: raw: Sequence of dataset paths (relative or absolute). Returns:

Normalise + validate an edit-method name (case-insensitive).

Word error rate — word-level edit distance / reference word count.

Return auth headers with the current UI token.

Return auth headers with the current UI token.

Load a covariance corpus (JSONL or plain text) for ``--cov-corpus`` (#250). Each JSONL object contributes its ``text`` / ``prompt`` / ``content``