-
Notifications
You must be signed in to change notification settings - Fork 0
All issues
Issue creation is restricted in this repository
Issues
is:issue state:open
is:issue state:open
Search results
Harden the scoring harness: switch-point metrics and a normalisation audit
evaluationEvaluation harness, metrics, benchmarksEvaluation harness, metrics, benchmarksStatus: Open.#11 In RahilOp/liquid-ai;Real-time streaming inference on edge devices
deploymentEdge, on-device, quantisation, streamingEdge, on-device, quantisation, streamingStatus: Open.#10 In RahilOp/liquid-ai;Extend to trilingual code-switching (JA-EN-X)
dataCorpora, augmentation, dataset constructionCorpora, augmentation, dataset constructionresearchExploratory direction, new scopeExploratory direction, new scopeStatus: Open.#9 In RahilOp/liquid-ai;Expand the augmentation pipeline beyond MUSAN noise
dataCorpora, augmentation, dataset constructionCorpora, augmentation, dataset constructionStatus: Open.#8 In RahilOp/liquid-ai;Mix real code-switching corpora into training, not just evaluation
dataCorpora, augmentation, dataset constructionCorpora, augmentation, dataset constructiontrainingFine-tuning recipes, ablations, hyperparametersFine-tuning recipes, ablations, hyperparametersStatus: Open.#7 In RahilOp/liquid-ai;Build a speaker-diverse real-speech evaluation set
evaluationEvaluation harness, metrics, benchmarksEvaluation harness, metrics, benchmarksresearchExploratory direction, new scopeExploratory direction, new scopeStatus: Open.#6 In RahilOp/liquid-ai;Version-control the real-speech predictions so Table 5 is reproducible
evaluationEvaluation harness, metrics, benchmarksEvaluation harness, metrics, benchmarkspaper-blockerMust be resolved before the paper can be submittedMust be resolved before the paper can be submittedStatus: Open.#5 In RahilOp/liquid-ai;Disentangle LoRA rank from data mixture in the final ablation step
paper-blockerMust be resolved before the paper can be submittedMust be resolved before the paper can be submittedtrainingFine-tuning recipes, ablations, hyperparametersFine-tuning recipes, ablations, hyperparametersStatus: Open.#4 In RahilOp/liquid-ai;Score the merged Whisper checkpoint on the FLEURS monolingual controls
evaluationEvaluation harness, metrics, benchmarksEvaluation harness, metrics, benchmarkspaper-blockerMust be resolved before the paper can be submittedMust be resolved before the paper can be submittedStatus: Open.#2 In RahilOp/liquid-ai;Recover the training recipe for the best LFM run (r32 encoder LoRA + FLEURS mix)
paper-blockerMust be resolved before the paper can be submittedMust be resolved before the paper can be submittedtrainingFine-tuning recipes, ablations, hyperparametersFine-tuning recipes, ablations, hyperparametersStatus: Open.#1 In RahilOp/liquid-ai;