The strongest result in the paper — LFM2.5-Audio + LoRA at 80.8% ScriptAcc / 27.0% MER (Table 5) — has no documented recipe. docs/asr-ft-training-log.md stops at Run 7 (lfm_lora_native_asr), which is a different configuration.
Right now that row is a measurement without a method, and it is the row the paper's constructive claim rests on.
What's needed
Recover and add to docs/asr-ft-training-log.md as a new run entry, matching the format of Runs 1–7:
- LoRA config: rank, alpha, dropout, exact target modules for LM and Conformer encoder
- Learning rates for both AdamW param groups, warmup, schedule
max_steps, batch size, and which checkpoint was selected- Data recipe: exact composition of the "FLEURS-balanced mix" — row counts per category, how much FLEURS was mixed in, and which split
- Validation loss curve, and the W&B run ID if it exists
Where to look
- W&B project
lfm25-asr-lora-jp (entity awsheshnath708-lfm) — run history should still have the config
checkpoints/ on the training box03-training/asr-ft/run_r32.sh looks like the closest candidate script
Done when
docs/asr-ft-training-log.md has a Run 8 entry, and paper/main.tex §6 no longer carries the red TODO block.
The strongest result in the paper — LFM2.5-Audio + LoRA at 80.8% ScriptAcc / 27.0% MER (Table 5) — has no documented recipe.
docs/asr-ft-training-log.mdstops at Run 7 (lfm_lora_native_asr), which is a different configuration.Right now that row is a measurement without a method, and it is the row the paper's constructive claim rests on.
What's needed
Recover and add to
docs/asr-ft-training-log.mdas a new run entry, matching the format of Runs 1–7:max_steps, batch size, and which checkpoint was selectedWhere to look
lfm25-asr-lora-jp(entityawsheshnath708-lfm) — run history should still have the configcheckpoints/on the training box03-training/asr-ft/run_r32.shlooks like the closest candidate scriptDone when
docs/asr-ft-training-log.mdhas a Run 8 entry, andpaper/main.tex§6 no longer carries the red TODO block.