Compute KL(p || q) over the union of token vocabularies. Laplace-smoothed: missing tokens in q get ``_EPS`` so divergence stays finite even o

Decide which adapter serves a request given its fingerprint key. Deterministic: the same ``(policy, request_key)`` always returns the same bu

Evolve ``seeds`` for ``rounds`` rounds; return valid ``EvolvedRow``s. Each round evolves every currently-live instruction; invalid evolutions

Run the preference-data linter over a sample of ``raw_rows``. ``fmt='auto'`` detects dpo vs. kto from the first row. Any other non-dpo/kto fo

Read JSONL, pick top-uncertainty rows, write out, return summary.

Pure-function scan over an in-memory weights map. Returns a ``ScanReport`` with per-layer findings and an overall verdict (``OK`` / ``WARN``

Mean repetition score across a batch of trajectories. Higher = more trajectory degeneration = closer to echo trap. Returns 0.0 on empty input

Wrap a reward fn to subtract a bounded penalty on a gamed proxy. The inner fn is called exactly once (verbatim) and its reward is reduced by

Whether HF's own gradient checkpointing should be switched on. ``StreamedDecoderLayer`` already wraps every layer in ``checkpoint(use_reentra

Split markdown ``text`` on ATX headings into section records. Each output record has three keys: ``section`` (heading text, or ``None`` for p

Strip reasoning-trace blocks from `text` per `parser`. Returns `text` unchanged when `parser` is `None`/empty, when the parser is unknown (de

Canonical kebab-case bank name (case-insensitive).

Validate ``budget`` is a positive int within sane bounds. Mirrors v0.41.0 / v0.62.0 numeric validator policy: bool-first rejection, non-int -

Validate a HuggingFace collection slug (``owner/slug-hash``).

Validate the codebook entry count. Bool-rejected (bool is a subclass of int), positive-int only, capped at :data:`MAX_CODEBOOK_SIZE` so a mis

Validate a base name; returns canonical entry from BUNDLED_PACKS.

Validate QAT configuration and return warnings/errors. Args: quantization: The quantization setting (4bit, 8bit, none). backend:

Validate a SAE repo id against the closed allowlist. Case-insensitive: returns the canonical (lower-case) entry that matches.

Per-user scaling vector — non-empty, finite, ≤MAX_VECTOR_DIM.

Validate a CheckList test kind. Case-insensitive.

Validate a list of trainer-plugin names. Returns canonical names.

Validate a projection dimensionality (vocab size). Bounds: ``[1, _MAX_VOCAB_SIZE=262144]``. Bool rejected per project bool-as-int policy.

Write rows as JSONL under cwd; returns realpath written. Atomic via staged-tempfile + ``os.replace`` (matches v0.43.0 Part D ``copy_bundle_to

Persist ``record`` under ``<base_dir>/<iteration_id>/iteration.json``.

Run a pre-resolved pattern against ``text``; hot-path inner loop. Centralised so the per-prompt scoring path in ``_refusal_rate`` avoids a re

Tiny module shaped like a decoder LM: embed + N layers + norm + head.

Compute leg-2 ``(base_scores, tuned_scores)`` maps over the general suite. Bundled suite names (v0.71.38 — the MCQ/arithmetic *and* the behaviour

Return a fresh base class for ``make_multipack_trainer_class`` whose instances expose ``train_dataset`` / ``data_collator`` / ``args`` and the

Prefix-sum LM with an explicit past_key_values cache. The last-token logits depend on the FULL prefix (cumulative embedding sum), so a correc

Adapter-switching fake with an explicit per-call KV cache. PREFIX-DEPENDENT logits (mirrors the minillm ``_CacheFakeLM`` prefix-sum design):

Build a fake Llama-shaped causal LM (nn.Module) for block-expansion tests.

Merge LoRA adapter with base model.

Parse a string value into the appropriate Python type.

Strip ANSI escape codes for keyword substring assertions.

Convert dataset rows to PPO format. PPO expects each row to have a 'prompt_text' field (string for tokenization) and optionally an 'answer' f

Resolve the checkpoint path from --resume argument. If resume == "auto", find the latest checkpoint in the output directory. Otherwise, treat

Return ``("local", path)`` or ``("hf", repo_id)`` for ``--model``. Falls through to HF when the local path doesn't exist *and* the value look

Measure one real step and let THAT decide, not the formula (#349). Raises when the measured peak does not fit. The streaming runtime is

Remove ANSI escape codes from Rich-formatted output.

A 2-layer Llama built from config — small enough for CPU CI, real enough that HF's decay/no-decay split behaves exactly as it does at 8B.

Write a list of dicts as JSONL.

Helper: write adapter_model.safetensors + minimal adapter_config.json.

Write a tiny PCM16 sine WAV using only the stdlib.

Assert ``load_config_from_string`` rejects ``yaml`` naming ``keyword``.

Quantise ``model`` with torchao's NVFP4 scheme (Blackwell-only). Live since v0.71.21 (#141). Routes through the same ``torchao.quantization.N

Freeze every parameter, then unfreeze those matching ``patterns``. This is the Spectrum (#266) targeted-training mechanism: full fine-tuning

Attach the v0.71.11 live RL callbacks; return how many were attached. Wires (when their schema fields are set): - reward-hacking detector (#2

v0.53.11 #68 — wrap inner trainer's ``compute_loss`` for TRUE weighted blend. Replaces the v0.40.1 primary-loss approximation with a per-batch

Return a ``CompilePrefixCallback`` for ``output_dir``.

Build a :class:`MultipackBatchSampler` from per-sample token lengths. Args: lengths: token counts for each sample in the (already tokeniz

Build the URL that a phone scans. Restrictions: - scheme must be 'http' or 'https' - plain http only allowed on loopback hosts (LAN expos

Compare two locks field-by-field. Returns a ``LockDrift`` report. Only the four content-bearing SHA fields + base_model name affect drift det

Split text into chunks no larger than ``max_chunk_chars``. Splits on paragraph boundaries first; falls back to hard slice when a paragraph al

Return a short tag describing the failure mode, or 'other'.

RLVR code reward: execute completion code, compare output to expected. Security: runs every completion in a subprocess sandbox with a 5s timeout

Standardized mean difference (pooled std) between two samples. Degenerate cases (fewer than 2 points in either sample, or zero pooled varianc

Aggregate entries by model, computing per-benchmark scores.

Build a deterministic cache key from the input tuple. Callers should pass FULL SHA-256 hex strings (64 chars) for the two digest inputs. The

Derive a calibrated linear probe from two labelled activation sets. ``positive`` / ``negative`` are ``[N, D]`` activation matrices (one row p

Relative loss change on A's domain when B is co-loaded. Returns ``(loss_a_with_b - loss_a_alone) / loss_a_alone``. Baseline loss must be > 0

Privacy score ∈ ``[0, 1]`` from membership-inference AUC. AUC ≈ 0.5 → no leak → score 1.0. AUC ≥ 0.9 or ≤ 0.1 → full leak → score 0.0. Symmet

Compute the aggregate echo signal from a buffer snapshot. Returns ``None`` when no completions are available.

Sparse-ReLU encoder: ``ReLU(activations @ w_enc + b_enc)``. Parameters ---------- activations : ``[N_tokens, D_model]`` numpy array of mo

Synthesise ``examples_per_endpoint`` rows for one endpoint. A row is one user-question / assistant-tool-call pair. We do NOT make network cal

Estimate model weight memory in GB. Args: model_params_b: Model size in billions of parameters. quantization: '4bit', '8bit', or

Export a BitNet model as a TQ1_0 (1.58-bit ternary) GGUF (v0.71.20 #134). Two-stage llama.cpp pipeline (reuses the v0.53.1 gguf machinery): 1

Flatten a chosen/rejected/completion field to plain text. DPO chosen/rejected may legitimately be message LISTS (conversational DPO) rather t

Pick the LR with the steepest negative gradient before divergence. Edge case: when the smoothed loss is monotonically increasing from the sta