Files
axolotl/docs/multimodal_assistant_mask.md
thad0ctor 5352d41d32 feat: systemic multimodal assistant-only loss masking + cfg.role_boundaries` (#3625)
* feat: systemic multimodal assistant-only loss masking + cfg.role_boundaries

Fixes silent ignoring of `cfg.train_on_inputs` / `cfg.roles_to_train` /
`cfg.train_on_eos` in the multimodal training path. Before this branch,
only Gemma 3n honored these knobs; every other VLM trained on the full
sequence regardless of config. Also adds `cfg.role_boundaries` YAML
override so users can declare per-role markers without subclassing.

What changed
------------
- `ProcessingStrategy` gains a declarative boundary scanner. Each
  strategy declares per-role start/end markers via
  `_build_role_boundaries`; the shared scanner honors
  `train_on_inputs` / `roles_to_train` / `train_on_eos` (incl. "last").
- New per-template strategies: Gemma 4, Llama 3.2 Vision, Llama 4,
  Pixtral, Mistral V7 Tekken.
- Refactored: Gemma 3 (previously no role masking), Gemma 3n
  (previously ad-hoc scanner, now shared).
- Strategies whose boundary tokens couldn't be verified offline
  (Voxtral, SmolVLM2, Mistral3, InternVL, GLM4V, llava/lfm2vl
  fallback) retain legacy behavior and emit a one-shot warning. Users
  can enable masking on them via `cfg.role_boundaries`.
- Pixtral / Mistral V7 Tekken correctly handle the shared `[/INST]`
  token between user-end and assistant-start via `include_end=False`
  + scanner rewind.

See `docs/multimodal_assistant_mask.md` for the full audit table,
root-cause analysis, and design rationale.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat: systemic multimodal assistant-only loss masking + cfg.role_boundaries

Fixes silent ignoring of `cfg.train_on_inputs` / `cfg.roles_to_train` /
`cfg.train_on_eos` in the multimodal training path. Before this branch,
only Gemma 3n honored these knobs; every other VLM trained on the full
sequence regardless of config. Also adds `cfg.role_boundaries` YAML
override so users can declare per-role markers without subclassing.

What changed
------------
- `ProcessingStrategy` gains a declarative boundary scanner. Each
  strategy declares per-role start/end markers via
  `_build_role_boundaries`; the shared scanner honors
  `train_on_inputs` / `roles_to_train` / `train_on_eos` (incl. "last").
- New per-template strategies: Gemma 4, Llama 3.2 Vision, Llama 4,
  Pixtral, Mistral V7 Tekken.
- Refactored: Gemma 3 (previously no role masking), Gemma 3n
  (previously ad-hoc scanner, now shared).
- Strategies whose boundary tokens couldn't be verified offline
  (Voxtral, SmolVLM2, Mistral3, InternVL, GLM4V, llava/lfm2vl
  fallback) retain legacy behavior and emit a one-shot warning. Users
  can enable masking on them via `cfg.role_boundaries`.
- Pixtral / Mistral V7 Tekken correctly handle the shared `[/INST]`
  token between user-end and assistant-start via `include_end=False`
  + scanner rewind.

See `docs/multimodal_assistant_mask.md` for the full audit table,
root-cause analysis, and design rationale.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* docs+types: address CodeRabbit nitpicks on PR #7

- builders/causal.py: add inline NOTE that multi-dataset configs reuse
  the first dataset's masking knobs (roles_to_train / train_on_eos) for
  all datasets — heterogeneous per-dataset overrides are not supported
  in the MM path today.
- processing_strategies.py: annotate inner scanner helpers
  _match_prefix and _find_end with explicit types (Tensor, int,
  list[int] → bool / tuple[int, bool]) for readability.
- docs/multimodal_assistant_mask.md: renumber the "Commits on this
  branch" list to 1-7 consecutive (previously skipped 3).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(mm-mask): address two CodeRabbit findings on PR #7

1. Schema rejected `train_on_eos: "none"` despite the scanner honoring it.
   `_VALID_TRAIN_ON_EOS` accepts "none" and the design doc lists it, but
   `SFTDataset.train_on_eos` was `Literal["all", "turn", "last"]`, so YAML
   users hit a pydantic ValidationError at config load. Added "none" to
   the Literal and updated the description.

2. `cfg.role_boundaries: []` had split-personality semantics: the strategy
   ctor treated it as "replace built-ins with empty" while the collator
   plumbing treated it as "unset", and both the design doc and the
   MultiModalConfig schema help text promised wholesale replacement for
   any set value. Aligned on opt-in semantics across all four surfaces —
   a non-empty list replaces built-ins wholesale; unset or `[]` falls back
   to built-ins. Rationale: honoring `[]` literally yields all-masked
   labels and zero gradient, which is almost always a typo or leftover
   rather than a deliberate user action. Users who want to disable role
   masking should unset the field or use `train_on_inputs: true`.

   Also sharpened the fallback one-shot warning for strategies without
   built-in boundaries: names the consequence ("only pad and media tokens
   are masked, every other token contributes to loss") and points users
   at `cfg.role_boundaries` + docs/multimodal_assistant_mask.md instead
   of "see axolotl/processing_strategies.py for how to declare
   boundaries."

Files:
- src/axolotl/utils/schemas/datasets.py: Literal adds "none"
- src/axolotl/processing_strategies.py: ctor truthiness check on
  role_boundaries_override; sharpened fallback warning
- src/axolotl/utils/schemas/multimodal.py: role_boundaries description
  now calls out opt-in + empty-list fallback semantics
- docs/multimodal_assistant_mask.md: same clarification in the Semantics
  block; updated the fallback-path detection paragraph to quote the new
  warning text
- tests/test_processing_strategies.py: +2 regressions
  (test_sft_dataset_schema_accepts_all_supported_train_on_eos_values,
  test_empty_role_boundaries_override_falls_back_to_builtin); 63/63 pass

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* doc cleanup

* fix(mm-mask): CodeRabbit findings + lint fix on PR #3625

Pre-commit failure: trailing newline missing on
docs/multimodal_assistant_mask.md (end-of-file-fixer hook).

Six CodeRabbit findings addressed:

1. Scanner: non-trainable role's end marker ignored ``include_end``.
   Under ``train_on_eos="all"``, the shared ``[/INST]`` token (user-end
   with ``include_end=False``, intentionally re-matched as assistant-start)
   leaked into loss via the user branch on Pixtral / Mistral V7 Tekken.
   Fix: gate the non-trainable branch on ``best_match.include_end`` to
   mirror the trainable branch.

2. Gemma3 ``boi_token`` lookup used ``tokenizer.special_tokens_map.get("boi_token")``,
   which never fires on real checkpoints (``special_tokens_map`` only
   holds HF's standard slots — bos/eos/pad/unk/...). Swap to direct
   attribute read ``getattr(tokenizer, "boi_token", None)``, matching
   what ``transformers.models.gemma3.processing_gemma3`` itself does.
   Updated the ``_gemma_tokenizer`` test fixture to mirror real-model
   shape so the test exercises the production code path.

3. GLM dispatcher only registered ``Glm46VProcessor`` (GLM-4.6V /
   GLM-4.7V). Real ``Glm4vProcessor`` (GLM-4V / GLM-4.1V) users fell
   through to the base fallback. Both processors ship identical
   media-token markers, so register both under the shared
   ``Glm4vProcessingStrategy`` with independent try/except import blocks.
   Updated class docstring. +2 dispatcher regressions.

4. Gemma3 ``process_labels`` hardcoded 262144 for the soft image token.
   Resolve dynamically via ``tokenizer.convert_tokens_to_ids("<image_soft_token>")``
   with unk-id guard; fall back to 262144 only if the string isn't in
   vocab. Mirrors ``Gemma4ProcessingStrategy.process_labels`` pattern.

5. ``build_collator`` was called twice per ``build()`` (eval + train
   passes), producing two identical ``MM collator: ...`` INFO banners on
   startup. Gate the log on ``is_eval=False`` so only the training pass
   emits it.

6. Removed unused ``_mistral_common_stub`` pytest fixture (13 refs → 0,
   always returned ``None``; the dispatcher already handles missing
   ``mistral_common`` via lazy import + ``try/except``). Added
   ``test_scanner_train_on_eos_all_with_non_trainable_include_end_false``
   — a focused scanner-level lock-in for finding #1, independent of any
   specific VLM strategy.

Test count: 63 → 68 passing. Local ``pre-commit run --all-files`` green.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* chore(mm-mask): hoist .tolist() out of scanner; shorten comments/docstrings

- Scanner perf: convert labels[i] to a Python list once per row so
  _match_prefix / _find_end operate on list slices instead of
  re-materializing Tensor slices via .tolist() on every probe. Cuts
  O(n*boundaries) CPython↔C boundary crossings per batch.
- Markdown lint (MD001, MD040): promote two h3 section headings to h2
  under the h1; add `text` language to the verify-at-runtime fenced block.
- Shorten verbose comments/docstrings added in recent commits to
  bare-minimum "why" notes matching the repo's existing style.

68/68 tests, 8/8 pre-commit hooks still pass.
2026-05-05 11:25:39 -04:00

3.2 KiB

Multimodal assistant-only loss masking

Correct placement

# Top-level: only train_on_inputs lives here.
train_on_inputs: false

datasets:
  - path: data/train.jsonl
    type: chat_template
    roles_to_train:          # per-dataset — this is what the MM scanner reads
      - assistant
    train_on_eos: turn       # per-dataset — same

test_datasets:
  - path: data/val.jsonl
    type: chat_template
    split: train
    roles_to_train:
      - assistant
    train_on_eos: turn

How to verify at runtime

build_collator logs the resolved knobs at INFO:

MM collator: train_on_inputs=False roles_to_train=['assistant'] train_on_eos=turn role_boundaries_override=none

If roles_to_train logs as None, the YAML knobs are not reaching the scanner — check that they are under datasets[0], not at the root.

Each verified strategy additionally logs its resolved boundary token ids at strategy init (e.g. <|turn>model[105, 4368], <turn|>[106] for Gemma 4). If a strategy emits the "has no built-in role boundaries ... only pad and media tokens are masked" one-shot warning instead, it is on the fallback path — declare per-role markers in YAML via cfg.role_boundaries (below) to activate masking. The strategies currently on this path are listed in the audit table above under fallback + warn.

Config-based override: cfg.role_boundaries

For the "unverified" strategies above, or for custom chat templates that don't match a built-in strategy's markers, users can declare role boundaries directly in YAML without subclassing:

role_boundaries:
  - role: assistant
    start: "<|turn>model"
    end: "<turn|>"
  - role: user
    start: "<|turn>user"
    end: "<turn|>"
  # Optional keys:
  # include_start: false   # default False
  # include_end: true      # default True, respects cfg.train_on_eos
  # end: eos_token         # sentinel: resolves to tokenizer.eos_token_id
  # end: null              # span runs to end of sequence

Semantics:

  • start and end are literal strings; axolotl encodes them at strategy init via tokenizer.encode(..., add_special_tokens=False) and logs the resolved token-id sequences at INFO level.
  • The special value end: eos_token is the portable way to express "Pixtral-style assistant turns end at EOS" without hard-coding an id.
  • role_boundaries is an opt-in override. A non-empty list replaces the strategy's built-in declarations wholesale (partial overlays are intentionally unsupported — they're hard to reason about at review time). Leaving the field unset or setting it to an empty list ([]) both mean "use the strategy's built-ins." Writing role_boundaries: [] is almost always a typo or leftover — honoring it literally would produce all-masked labels and zero gradient, so it is treated the same as unset.
  • cfg.roles_to_train still governs which declared roles contribute to loss. You can declare user and assistant boundaries and set roles_to_train: ["assistant"] to have the scanner correctly identify user spans as masking boundaries without training on their content.
  • Invalid specs fail loudly at strategy init (missing role/start, unencodable markers), not silently at loss-compute time.