How franzerz/aba-mining-tasks-llms-group-kmitl-internship-2026 works
This repository is an experimental ABA-mining toolkit. Its main runtime workflow runs three LLM-assisted tasks: Task 1 extracts topic, span, and sentiment annotations; Task 2 generates body literals from gold or Task 1 outputs; Task 3 classifies contrary relations between phrase pairs. Separate evaluator scripts assess task outputs, while finetune scripts support Task 1 model experiments. The README gives little product detail, so these capabilities are inferred from the source tree and sampled task code; most evaluator and finetune internals were not inspected.
Main entry points: llm.py LLM client · config.py Runtime configuration · task2.py Body literals · task1.py Topic span sentiment
Components 20
Task runners
- Task 1 runner
ABA_mining/run_task1.py - Task 2 runner
ABA_mining/run_task2.py - Task 3 runner
ABA_mining/run_task3.py
Mining pipeline
- Topic span sentiment
ABA_mining/src/task1.py - Body literals
ABA_mining/src/task2.py - Contrary classification
ABA_mining/src/task3.py - Prompt assembly
ABA_mining/src/prompts.py - LLM client
ABA_mining/src/llm.py - Runtime configuration
ABA_mining/src/config.py - JSON normalization
ABA_mining/src/utils.py
Evaluation
- Task evaluators
ABA_mining/evaluator - Result summaries
ABA_mining/results.py - Task 3 candidates
ABA_mining/build_task3_candidates.py - Dataset preparation
ABA_mining/prepare_dataset.py
Task 1 finetuning
- Training data prep
ABA_mining/finetune/prepare_task1_data.py - LoRA training
ABA_mining/finetune/train_task1_lora.py - Finetuned inference
ABA_mining/finetune/run_task1_finetuned.py - Finetune evaluation
ABA_mining/finetune/eval_task1_2_ft.py
Other components
- Experiment operator
- LLM providers
Connections 29
- Experiment operator → Task 1 runner: starts
- Experiment operator → Task 2 runner: starts
- Experiment operator → Task 3 runner: starts
- Task 1 runner → Prompt assembly: assembles prompt
ABA_mining/run_task1.py - Task 1 runner → LLM client: imports client
ABA_mining/run_task1.py - Task 2 runner → Body literals: runs
ABA_mining/run_task2.py - Task 2 runner → LLM client: imports client
ABA_mining/run_task2.py - Task 3 runner → Contrary classification: runs
- Topic span sentiment → Runtime configuration: uses
ABA_mining/src/task1.py - Topic span sentiment → LLM client: uses
ABA_mining/src/task1.py - Topic span sentiment → Prompt assembly: uses
ABA_mining/src/task1.py - Topic span sentiment → JSON normalization: uses
ABA_mining/src/task1.py - Body literals → Runtime configuration: uses
ABA_mining/src/task2.py - Body literals → LLM client: uses
ABA_mining/src/task2.py - Body literals → Prompt assembly: uses
ABA_mining/src/task2.py - Body literals → JSON normalization: uses
ABA_mining/src/task2.py - Contrary classification → Runtime configuration: uses
ABA_mining/src/task3.py - Contrary classification → LLM client: uses
ABA_mining/src/task3.py - Contrary classification → Prompt assembly: uses
ABA_mining/src/task3.py - LLM client → LLM providers: calls provider
ABA_mining/src/llm.py - Task 1 runner → Runtime configuration: loads settings
- Task 2 runner → Runtime configuration: loads settings
ABA_mining/run_task2.py - Task 3 runner → Runtime configuration: loads settings
- Task evaluators → Result summaries: summarizes outputs
- Task 3 candidates → Contrary classification: prepares inputs
- Dataset preparation → Topic span sentiment: prepares data
- Training data prep → LoRA training: prepares training data
- LoRA training → Finetuned inference: trains model for
- Finetuned inference → Finetune evaluation: evaluated by