How xlang-ai/unifiedskg works
UnifiedSKG is a multi-task structured-knowledge-grounding research framework: it converts heterogeneous tasks to text-to-text datasets, trains or evaluates sequence-to-sequence models, and scores generated results. The graph highlights the documented task breadth and model-tuning variants alongside the training path. Some connections are documented only at a high level; source excerpts are partial, so uncertain runtime links are omitted or marked unverified.
Main entry points: train.py Training entry · metrics Task evaluators · modeling_auto.py Prompt model selection · seq2seq_construction Seq2seq construction
Components 24
Training workflow
- Training entry
train.py
Task data
- Task definitions
tasks - Dataset loading
utils/dataset.py - Seq2seq construction
seq2seq_construction - Dialogue expansion
seq2seq_construction/multi_woz_22_response.py - Table linearization
utils/processor/table_linearize.py - Table truncation
utils/processor/table_truncate.py
Model variants
- Model base
models/unified/base.py - Prompt models
models/prompt/modeling_t5.py - Prompt model selection
models/prompt/modeling_auto.py - BART prompt model
models/prompt/modeling_bart.py - Adapter T5
models/adapter/modeling_t5.py - Adapter layers
models/adapter/adapter.py - Multi-task prefix tuning
models/unified/combined_prefixtuning.py
Training runtime
- Configuration
utils/configue.py - Seq2seq trainer
utils/trainer.py - Training arguments
utils/training_arguments.py - Runtime utilities
utils/tool.py
Evaluation
- Task evaluators
metrics - Unified exact match
metrics/unified/evaluator.py - MTOP metrics
metrics/mtop/top_metrics.py - WikiSQL query parsing
metrics/wikisql_fully_supervised/query.py - WikiSQL execution
metrics/wikisql_fully_supervised/dbengine.py
Other components
- Researcher
Connections 21
- Researcher → Training entry: starts training
README.md - Training entry → Configuration: loads settings
train.py - Training entry → Dataset loading: loads datasets
train.py - Training entry → Seq2seq construction: builds seq2seq data
train.py - Training entry → Runtime utilities: resolves constructor
train.py - Training entry → Training arguments: uses arguments
train.py - Training entry → Seq2seq trainer: runs training
train.py - Task definitions → Seq2seq construction: task coverage
README.md - Seq2seq construction → Dialogue expansion: includes dialogue data
seq2seq_construction/multi_woz_22_response.py - Table truncation → Table linearization: uses linearizer
utils/processor/table_truncate.py - Seq2seq trainer → Task evaluators: evaluates predictions
- Task evaluators → Unified exact match: includes shared scoring
README.md - Task evaluators → MTOP metrics: includes task scoring
metrics/mtop/evaluator.py - MTOP metrics → Task evaluators: uses task utilities
metrics/mtop/top_metrics.py - Task evaluators → WikiSQL query parsing: includes SQL parsing
metrics/wikisql_fully_supervised/dbengine.py - WikiSQL execution → WikiSQL query parsing: executes parsed query
metrics/wikisql_fully_supervised/dbengine.py - Prompt model selection → Prompt models: selects model family
models/prompt/modeling_auto.py - Prompt model selection → BART prompt model: selects BART
models/prompt/modeling_auto.py - Adapter T5 → Adapter layers: uses adapter layers
models/adapter/modeling_t5.py - Multi-task prefix tuning → Model base: extends base model
models/unified/combined_prefixtuning.py - Multi-task prefix tuning → Prompt model selection: uses prompt model classes
models/unified/combined_prefixtuning.py