TI Coding Agent Data Prep Skill

Use this repo for Open-SWE-Traces data preparation before coding-agent SFT experiments.

Purpose

This project organizes the existing Open-SWE-Traces probing, native-trace filtering, and trajectory repurposing scripts into a Python project. It is intended for work on nvidia/Open-SWE-Traces, especially:

inspecting dataset structure and source splits,
auditing trajectories for SFT-quality hard filters,
measuring Qwen tokenizer/token statistics,
decomposing trajectories into task-oriented phases,
exporting balanced ModelScope-SWIFT training/validation probes,
experimenting with pi-mono-style conversion.

Repository Layout

scripts/probing/: dataset inspection, unique-count reports, tokenizer comparison, and token counting.
scripts/filtering/: native trajectory audit and hard filtering.
scripts/repurposing/: subproblem decomposition, coarse task-stage conversion, SWIFT export, and pi-mono conversion.
src/ti_coding_agent_data_prep/openswe/: shared constants and path helpers.
data/: local datasets; ignored by git.
runs/: script outputs; ignored by git.
docs/: retained legacy notes from the original probe workspace.

Expected Data Location

By default scripts read:

data/Open-SWE-Traces

On B300, prefer a symlink instead of copying the dataset:

ln -s /ssd/workspace/yi/openswetraces_probe/Open-SWE-Traces data/Open-SWE-Traces

Environment

Use a repo-local environment. On B300, set proxy variables before installing or downloading:

export http_proxy=http://100.72.0.101:8888
export https_proxy=http://100.72.0.101:8888
export HTTP_PROXY=http://100.72.0.101:8888
export HTTPS_PROXY=http://100.72.0.101:8888
export HF_ENDPOINT=https://hf-mirror.com
uv venv .venv --python 3.10
source .venv/bin/activate
uv pip install -e '.[dev]'

Main Entrypoints

Probe one sample per split:

python scripts/probing/inspect_sample.py

Generate unique-count and random-20 decomposition reports:

python scripts/probing/analyze_and_decompose.py

Run native hard-filter audit:

python scripts/filtering/audit_native_traces.py --input-root data/Open-SWE-Traces --output-dir runs/native_trace_audit

Run exact Qwen token counting:

python scripts/probing/count_qwen_tokens_exact_parallel.py --input-root data/Open-SWE-Traces --output runs/native_trace_audit/qwen_exact_token_count.json --model Qwen/Qwen3-32B --workers 12

Build the balanced 5k SWIFT training probe:

python scripts/repurposing/build_swift_training_probe_5k.py

Build the balanced 500-row validation probe:

python scripts/repurposing/build_swift_validation_500.py

Try pi-mono-style conversion:

python scripts/repurposing/convert_openswe_to_pi_mono.py --input-root data/Open-SWE-Traces --output-root runs/pi_mono_converted

Filtering Semantics

The hard filter rejects trajectories with malformed tool-call JSON, tool-call/tool-result mismatch, polluted tool names or arguments, empty or inconsistent final patches, excessive length, repeated no-progress loops, and unresolved off-track traces.

Training Format Notes

For SWIFT exports:

MiniMax trajectories are treated as thinking-mode data and reasoning_content is wrapped with <think>...</think>.
Qwen trajectories are treated as non-thinking-mode data.
system, user, and tool messages use loss=false; assistant messages use loss=true.

3.4 KiB Raw Blame History