Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@acestep/ui/gradio/i18n/he.json`:
- Around line 447-475: Update the Hebrew locale file to translate the remaining
English LoKr strings: replace values for keys train_section_tensors,
train_section_lora, train_section_params, lokr_section_tensors,
lokr_section_settings, lokr_linear_dim, lokr_linear_dim_info, lokr_linear_alpha,
lokr_linear_alpha_info, lokr_factor, lokr_factor_info, lokr_decompose_both,
lokr_decompose_both_info, lokr_use_tucker, lokr_use_tucker_info,
lokr_use_scalar, lokr_use_scalar_info, lokr_weight_decompose,
lokr_weight_decompose_info, lokr_learning_rate_info, lokr_checkpoint_epoch, and
lokr_checkpoint_epoch_info with appropriate Hebrew translations consistent with
existing style (keep code/backtick examples like `.pt` as-is and preserve
emoji/button text where present).

In `@acestep/ui/gradio/i18n/ja.json`:
- Around line 434-456: Several LoKr locale entries are still in English; update
the specified keys (lokr_linear_dim_info, lokr_linear_alpha_info,
lokr_factor_info, lokr_decompose_both_info, lokr_use_tucker_info,
lokr_use_scalar_info, lokr_weight_decompose_info, lokr_learning_rate_info,
lokr_checkpoint_epoch, lokr_checkpoint_epoch_info) to proper Japanese
translations consistent with surrounding entries (e.g., explain rank/dimension,
scaling factor, Kronecker factor, decomposing both sides, Tucker decomposition,
scalar gating, weight decomposition, learning rate guidance, checkpoint epoch
label and selection guidance) and replace the English strings in the ja.json
file accordingly.

In `@acestep/ui/gradio/i18n/zh.json`:
- Around line 434-456: Several LoKr localization strings in zh.json are still in
English; update the listed keys (lokr_linear_dim_info, lokr_linear_alpha_info,
lokr_factor_info, lokr_decompose_both_info, lokr_use_tucker_info,
lokr_use_scalar_info, lokr_weight_decompose_info, lokr_learning_rate_info,
lokr_checkpoint_epoch, lokr_checkpoint_epoch_info) with proper Chinese
translations consistent with the surrounding entries, preserving meaning (e.g.,
describe rank/dimension, scaling factor, Kronecker factor, decomposing both
sides, Tucker decomposition, scalar gating, weight decomposition, learning rate
guidance, and checkpoint selection) and keep any UI emojis or punctuation style
consistent with other zh.json entries.

In `@acestep/ui/gradio/interfaces/training_dataset_tab_label_preview.py`:
- Around line 70-95: The hardcoded English UI strings in the training dataset
preview need localization: update the Textbox and Dropdown initializers
(edit_caption, edit_genre, prompt_override, edit_lyrics) to use the translation
function t(...) for placeholders, the prompt_override choices/values, and any
labels currently using literals; specifically replace "Music description...",
"pop, electronic, dance...", the Dropdown choices/value ("Use Global Ratio",
"Caption", "Genre"), and the lyrics placeholder "[Verse 1]\nLyrics
here...\n\n[Chorus]\n..." with appropriate t(...) keys (e.g.,
t("training.caption_placeholder"), t("training.genre_placeholder"),
t("training.prompt_override_use_global"), t("training.prompt_override_caption"),
t("training.prompt_override_genre"), t("training.lyrics_placeholder")) so all
strings are localized.

In `@acestep/ui/gradio/interfaces/training_dataset_tab_save_preprocess.py`:
- Around line 64-70: The Dropdown instantiation for preprocess_mode hardcodes
English strings; wrap the label, choices and info text in the translation
function (t) so they render localized UI copy. Update the gr.Dropdown call for
preprocess_mode to use t("...") for label ("Preprocess For"), each choice
("LoRA", "LoKr") via translated strings or a mapped list, and the info string
("LoRA keeps compatibility mode; LoKr uses per-sample source-style context.") so
all displayed text uses t(...) instead of literal English.

In `@acestep/ui/gradio/interfaces/training_dataset_tab_scan_settings.py`:
- Around line 13-40: The UI contains hardcoded English in the quick-start
paragraph and two section headers; replace those literal strings with
translation lookups (use t("training.quick_start_message") for the paragraph and
t("training.load_section_title") and t("training.scan_section_title") for the
headings) inside the existing gr.HTML and gr.HTML("<h4>...") calls (referenced
as the first gr.HTML block and the two gr.HTML header calls) and add
corresponding keys to the locale resource files so all locales render the same
copy.
- Around line 59-95: The Dataframe headers and the LM checkbox label/info are
hardcoded English; update audio_files_table to use translated strings via the
t(...) function for each header (e.g., replace "#", "Filename", "Duration",
"Lyrics", "Labeled", "BPM", "Key", "Caption" with t(...) keys you add like
training.table_index, training.table_filename, etc.) and replace the static
label/info for format_lyrics and transcribe_lyrics with t(...) keys (e.g.,
training.format_lyrics_label, training.format_lyrics_info,
training.transcribe_lyrics_label, training.transcribe_lyrics_info) so all
user-facing text in audio_files_table, format_lyrics, and transcribe_lyrics is
localized.

In `@acestep/ui/gradio/interfaces/training_lora_tab_run_export.py`:
- Around line 91-95: The UI strings for the Textbox named resume_checkpoint_dir
are not localized; update the Textbox initialization so its label and info use
the localization function (e.g., t("...")) instead of hardcoded English text —
replace label="Resume Checkpoint" with label=t("...") and info="Directory of a
saved LoRA checkpoint to resume from" with info=t("...") in the
resume_checkpoint_dir declaration inside training_lora_tab_run_export.py to
ensure locale consistency.

In `@acestep/ui/gradio/interfaces/training.py`:
- Around line 40-47: The HTML header currently hardcodes English inside the
gr.HTML(...) call; replace the static English strings with localized text by
calling the localization function (e.g., t("...")) for both the title and
subtitle and injecting those localized values into the HTML template used in
gr.HTML. Locate the gr.HTML(...) block in training.py and change the embedded
"<h2>🎵 LoRA Training for ACE-Step</h2>" and the <p> subtitle to use t("...")
results (for example t("training.header.title") and
t("training.header.subtitle")) so the rendered UI uses translated text.

---

Nitpick comments:
In `@acestep/ui/gradio/interfaces/training_dataset_tab_label_preview.py`:
- Around line 10-11: Update the docstring for
build_dataset_label_and_preview_controls to include the return contract: list
the dict keys returned (e.g., keys for auto-label controls, sample-preview
components, any Gradio Blocks/Elements) and describe the type/value for each key
and their role in the wiring/facade (for example "auto_label_controls: Gradio
Block containing inputs/buttons", "preview_panel: Gradio Component showing
sample preview", etc.), and mention any exceptions or error conditions; keep it
concise and follow existing module docstring style.

In `@acestep/ui/gradio/interfaces/training.py`:
- Around line 38-59: The epoch defaults resolved by
_resolve_epoch_slider_defaults() are not passed to create_training_lokr_tab(),
so LoKr uses hardcoded epochs; update the call in create_training_lokr_tab()
invocation to forward epoch_min, epoch_step, epoch_default, and then modify the
create_training_lokr_tab function signature in both training_lokr_tab.py and
training_lokr_tab_run_export.py to accept (epoch_min, epoch_step, epoch_default)
and use those values for the LoKr epoch slider (and any defaults/exports)
instead of the fixed constants so debug/default parity is maintained.