MOSS training: request example dataset files for instruction (emotion) field

[krisheeetu]

I've read MOSS_TRAINING_GUIDE.md and MOSS_TTS_PROMPT_FIELDS_GUIDE.md, but I'm still unsure how to correctly format the instruction field for emotion-conditioned training on laion/moss-tts-v1.5-8b-voice-acting.

Could you share a small example dataset (a few rows of manifest.jsonl, or example instruction_lines values) showing a working instruction-conditioned training setup?

Also, does the folder auto-import on MOSS Dataset Prep support instruction at all,
or is a JSONL manifest via MOSS Dataset Rows required to use it?

A minimal sample (3-5 rows) in the docs or example_workflows would help a lot. Thanks!

[diodiogod]

I only implemented the MOSS training integration; I did not create the upstream training method or the LAION Voice Acting fine-tune. My own test was a LoRA for one specific voice, without instructions, so I cannot claim a tested best practice for instruction-conditioned training.

From checking the code:

• instruction is preserved in the JSONL and passed directly to the official MOSS processor.build_user_message(...) for each training row.
• MOSS Dataset Rows supports it through instruction_lines: one line per clip, in the same order as text_lines.
• Folder auto-import does not support instructions. It only converts matching clip.wav + clip.txt files into audio + text rows. To use per-clip instructions, use MOSS Dataset Rows or provide your own JSONL manifest.

Structurally, a manifest can look like this:

{"audio":"./clips/001.wav","text":"I cannot believe you did that.","instruction":"Angry and confrontational, speaking loudly.","language":"English"}
{"audio":"./clips/002.wav","text":"Everything is going to be all right.","instruction":"Warm, gentle, and reassuring.","language":"English"}
{"audio":"./clips/003.wav","text":"Did you hear something downstairs?","instruction":"Fearful whisper, tense and hesitant.","language":"English"}

With MOSS Dataset Rows, put the three transcripts in text_lines and the three matching descriptions in instruction_lines.

However, these rows demonstrate the supported format, not a recipe I have personally validated. The official OpenMOSS trainer maps instruction into the prompt in the same way, but its clearest text + instruction training examples are for VoiceGenerator rather than a documented emotion-training recipe for MOSS-TTS v1.5. The LAION model card says performance direction, emotion, character, pacing, and similar cues belong in instruction, but it does not publish the actual training manifest or label format used for that checkpoint.

So you will ultimately need to experiment with wording and dataset consistency. I would start with short, consistent natural-language descriptions, use the same vocabulary for recurring emotions/styles, set language, and compare a small LoRA trained with instructions against a plain audio + transcript baseline before committing to a large run.