PIPELINE_USAGE.md

CodeFuse-DeBench Pipeline Usage

This document explains how to run the full pipeline for a single task.

1. Entry Script

Primary single-task entry:

• scripts/run_single_task.py

Execution boundary:

1. Host Step1: Readability
2. Guest Step2: Syntactic Correctness / Recompilation
3. Guest Step3: Semantic Fidelity

2. Basic Command

python3 scripts/run_single_task.py \
  src/7.c \
  decompiled/retdec_out/arm32/7/7_gcc_O2_no_g.c \
  --arch arm32 \
  --original-bin build/arm32/7/7_gcc_O2_no_g \
  --results-dir runs/qwen_demo \
  --llm-profile qwen3.5-plus

Argument summary:

• First positional argument: original source file
• Second positional argument: decompiled source file
• --arch: target architecture used to select the Lima instance
• --original-bin: original binary
• --results-dir: root directory of the target results tree
• --llm-profile: active profile name

3. Common Flags

--restart

Force a fresh rerun of the task instead of reusing existing Step2 outputs.

--skip-readability

Skip host-side Step1. Use only for debugging.

--skip-step3

Run Step1 + Step2 only and skip Step3.

--step3-only

Require an existing successful Step2 result and rerun only Step3.

4. Output Directory

Outputs are written to:

results_<llm>_v4_full/<arch>/<src>/<bin_name>/<decompiler>/

Each task directory contains:

• readability/
• syntactic/
• semantic/
• report.md

5. Result Interpretation

• readability/test_results_*.json: structured Step1 output
• syntactic/repair_trace.json: iterative Step2 repair trace
• semantic/result_metrics.json: Step3 metric summary
• report.md: per-task report

6. Environment Requirements

• The corresponding Lima instance is running
• If guest time drifts, synchronize the guest clock first
• Profiles and key inventory are configured under config/
• Run scripts/setup_lima_*.sh when guest setup is required