omkarmulakala/curio_next

How omkarmulakala/curio_next works

Audio Context Layer adapts Qwen2-Audio-7B-Instruct with LoRA for multiple-choice, audio-grounded question answering. Dataset preparation builds and validates leakage-remediated splits; training loads records, prepares audio/question/answer batches, adapts the language-model decoder, and saves an adapter. For use, inference combines an audio recording, question, and choices to generate an answer; evaluation compares predictions with references and reports metrics and errors. Reporting and training-curve plotting are additional analysis capabilities. CLI wiring is described in the README, but several scripts were not sampled, so uncertain internal calls are omitted.

Main entry points: run_inference.py Inference command · train.py Train command · evaluate.py Evaluation metrics · dataset_loader.py Manifest loader

Components 27

Dataset preparation

Prepare manifests scripts/prepare_dataset.py
Build manifests scripts/build_manifests.py
Split and leakage checks src/data/splitting.py
Manifest loader src/data/dataset_loader.py
Dataset validation src/data/validation.py

Model and training

Train command scripts/train.py
Conversation preparation src/data/preprocessing.py
Batch collator src/data/collator.py
Audio loading src/utils/audio.py
Qwen2-Audio src/models/qwen2_audio.py
LoRA adaptation src/models/lora.py
Trainer src/training/trainer.py

Answering

Inference command scripts/run_inference.py
Answer generation src/models/qwen2_audio.py

Evaluation and analysis

Evaluation command scripts/evaluate.py
Evaluation metrics src/evaluation/evaluate.py
Answer matching src/evaluation/metrics.py
Error analysis src/evaluation/error_analysis.py
Report generation scripts/generate_report.py
Training plots src/training/plots.py

Other components

Researcher
DCASE data
Audio recordings
Audio QA inputs
Saved adapter
Predictions
Metrics and reports

Connections 35

Researcher → Prepare manifests: runs preparation
DCASE data → Prepare manifests: provides metadata
Prepare manifests → Build manifests: builds manifests
Build manifests → Split and leakage checks: assigns splits
Split and leakage checks → Manifest loader: defines split records
Researcher → Dataset validation: requests validation
Dataset validation → Manifest loader: loads splits
Researcher → Train command: starts training
Train command → Manifest loader: loads records
Train command → Qwen2-Audio: loads model
Train command → LoRA adaptation: applies adapters
Train command → Trainer: builds trainer
Trainer → Batch collator: batches records
Batch collator → Conversation preparation: formats prompts
Batch collator → Audio loading: loads audio
Trainer → Saved adapter: saves adapter
Researcher → Inference command: requests answers
Inference command → Manifest loader: loads records
Inference command → Audio loading: decodes audio
Inference command → Qwen2-Audio: loads model
Inference command → Saved adapter: loads adapter
Inference command → Answer generation: requests generation
Answer generation → Conversation preparation: formats prompt
Audio QA inputs → Inference command: supplies questions
Audio recordings → Inference command: supplies recordings
Answer generation → Predictions: produces answers
Researcher → Evaluation command: requests evaluation
Predictions → Evaluation command: provides predictions
Evaluation command → Evaluation metrics: evaluates records
Evaluation metrics → Answer matching: computes accuracy
Evaluation metrics → Error analysis: identifies errors
Evaluation metrics → Metrics and reports: writes metrics
Researcher → Report generation: requests report
Report generation → Metrics and reports: generates report
Researcher → Training plots: requests plots

Drawn by GitDiagram on September 28, 2026 from the repository's file tree, README and source files. Read as Markdown ·