feat: systemic multimodal assistant-only loss masking + #3617#4

Closed

thad0ctor wants to merge 24 commits into

mainfrom

feat/multimodal-assistant-mask-all

Commits on Apr 23, 2026

feat: systemic multimodal assistant-only loss masking

thad0ctor
and
claude
committed
feat: thread cfg.train_on_inputs / roles_to_train / train_on_eos into MM collator

thad0ctor
and
claude
committed
feat: forward cfg.processor_kwargs to processor from_pretrained (#3617)

thad0ctor
and
claude
committed
docs: multimodal assistant-mask design doc

thad0ctor
and
claude
committed
feat: cfg.role_boundaries YAML override for MM role-mask scanner

thad0ctor
and
claude
committed
test: additional coverage for MM role-mask scanner edge cases

thad0ctor
and
claude
committed
chore: tighten docstrings and comments in multimodal mask refactor

thad0ctor
and
claude
committed
fix: resolve MM per-dataset masking knobs for pydantic SFTDataset

thad0ctor
and
claude
committed
fix(processing_strategies): correctly mask Pixtral / Mistral V7 Tekken assistant spans

thad0ctor
and
claude
committed
docs: refresh test count and commit list in MM mask design doc

thad0ctor
and
claude
committed
feat(processing_strategies): support train_on_eos='last' and validate at init

thad0ctor
and
claude
committed
fix(processing_strategies): honor explicit empty roles_to_train list

thad0ctor
and
claude
committed
fix(processing_strategies): guard None tokenizer id comparisons

thad0ctor
and
claude
committed
fix(processing_strategies): raise on unencodable end markers in role_boundaries

thad0ctor
and
claude
committed
fix(processing_strategies): narrow optional-import exception handling + debug log

thad0ctor
and
claude
committed
fix(processing_strategies): drop unreachable qwen3_5_moe dispatcher branch

thad0ctor
and
claude
committed
fix(processing_strategies): add trailing newline to Gemma4 role-start markers

thad0ctor
and
claude
committed
fix(processing_strategies): drop unreachable Gemma system boundary

thad0ctor
and
claude
committed
feat(processing_strategies): INFO-log resolved masking config at init

thad0ctor
and
claude
committed
docs: reflect Pixtral scanner fix + final test count in audit table

thad0ctor
and
claude
committed
revert: drop cfg.processor_kwargs / #3617 — tracked on feat/processor-kwargs

thad0ctor
and
claude
committed
fix(processing_strategies): clear three mypy errors flagged by CodeRabbit

thad0ctor
and
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: systemic multimodal assistant-only loss masking + #3617#4

feat: systemic multimodal assistant-only loss masking + #3617#4
thad0ctor wants to merge 24 commits into
mainfrom
feat/multimodal-assistant-mask-all

Commits on Apr 23, 2026

feat: systemic multimodal assistant-only loss masking

feat: thread cfg.train_on_inputs / roles_to_train / train_on_eos into MM collator

feat: forward cfg.processor_kwargs to processor from_pretrained (#3617)

docs: multimodal assistant-mask design doc

feat: cfg.role_boundaries YAML override for MM role-mask scanner

test: additional coverage for MM role-mask scanner edge cases

chore: tighten docstrings and comments in multimodal mask refactor

fix: resolve MM per-dataset masking knobs for pydantic SFTDataset

fix(processing_strategies): correctly mask Pixtral / Mistral V7 Tekken assistant spans

docs: refresh test count and commit list in MM mask design doc

feat(processing_strategies): support train_on_eos='last' and validate at init

fix(processing_strategies): honor explicit empty roles_to_train list

fix(processing_strategies): guard None tokenizer id comparisons

fix(processing_strategies): raise on unencodable end markers in role_boundaries

fix(processing_strategies): narrow optional-import exception handling + debug log

fix(processing_strategies): drop unreachable qwen3_5_moe dispatcher branch

fix(processing_strategies): add trailing newline to Gemma4 role-start markers

fix(processing_strategies): drop unreachable Gemma system boundary

feat(processing_strategies): INFO-log resolved masking config at init

docs: reflect Pixtral scanner fix + final test count in audit table

revert: drop cfg.processor_kwargs / #3617 — tracked on feat/processor-kwargs

fix(processing_strategies): clear three mypy errors flagged by CodeRabbit

Commits on Apr 24, 2026

fix(multimodal-cpt): resolve CI pre-commit + caplog flakiness on PR #4

fix(processing_strategies): address CodeRabbit findings on messages/vocab handling