DebugBelief · Pipeline

SWE-bench

GitHub Search

GitHub Expansion

GPU / HLM

—

SWE Tier 1 Queue

score ≥ 0.310

—

GitHub Collector

—

LLM Extractor

—

Traj Builder

5 gates

Quality Filter

llm≥0.35 · corrections

—

Gold

—

Silver

—

SWE Tier 2 Queue

score 0.220–0.305

—

GitHub Collector

—

LLM Extractor

—

Traj Builder

5 gates

Quality Filter

llm≥0.35 · corrections

—

Gold

—

Silver

—

GitHub Tier 1 Queue

score ≥ pool median

—

GitHub Collector

—

LLM Extractor

—

Traj Builder

5 gates

Quality Filter

pre-filtered during batch

—

Gold

—

Silver

—

GitHub Tier 2 Queue

score < pool median

—

GitHub Collector

—

LLM Extractor

—

Traj Builder

5 gates

Quality Filter

pre-filtered during batch

—

Gold

—

Silver

—

Exp A – Python

score ≥ mean · python repos

—

LLM Extractor

collect_expansion.py

—

Traj Builder

5 gates

Quality Filter

llm≥0.35 · corrections

—

Gold

—

Silver

—

Exp A – No-Python

TS · Go · Rust · k8s · Next.js

—

LLM Extractor

collect_expansion.py

—

Traj Builder

5 gates

Quality Filter

7 language sub-batches

—

Gold

—

Silver

—

Exp B – JVM + Tier C

spring · kafka · rails · node

—

LLM Extractor

collect_expansion.py

—

Traj Builder

5 gates

Quality Filter

pre-filtered during batch

—

Gold

—

Silver

—

Exp B – Mix

diverse repo search

—

LLM Extractor

collect_expansion.py

—

Traj Builder

5 gates

Quality Filter

pre-filtered during batch

—

Gold

—

Silver

—

Exp B – Python

python repos · running

—

LLM Extractor

collect_expansion.py

—

Traj Builder

5 gates

Quality Filter

pre-filtered during batch

—

Gold

—

Silver

—

Exp B – Rust+Go

rust · golang repos · running

—

LLM Extractor

collect_expansion.py

—

Traj Builder

5 gates

Quality Filter

pre-filtered during batch

—

Gold

—

Silver

—

GPU Libs Search

cupy · numba · xla · DeepSpeed

—

LLM Extractor

collect_expansion.py

—

Traj Builder

5 gates

Quality Filter

pre-filtered during batch

—

Gold

—

Silver

—

GPU Inference Search

vllm · onnxruntime

—

LLM Extractor

collect_expansion.py

—

Traj Builder

5 gates

Quality Filter

pre-filtered during batch

—

Gold

—

Silver

—

GPU Compilers Search

triton · tvm · jax

—

LLM Extractor

collect_expansion.py

—

Traj Builder

5 gates

Quality Filter

pre-filtered during batch

—

Gold

—

Silver

—

Gold Final

gold/gold_final.jsonl

—

Silver Final

processed/silver_final.jsonl

—

Gold Extended

gold/gold_extended.jsonl

—

Gold v2

gold/gold_v2.jsonl

—

Dev Split

splits/dev.jsonl

—

Test Split

splits/test.jsonl

—

Few-shot Pool

splits/few_shot_pool.jsonl

Input Queue

Processing

Gold

Silver

Output Split

Click any node for details

Select a node to view
statistics and entries.