jiliao's picture YipengGao's picture
Duplicate from YipengGao/3DCodeBench
6ad95bc
|
Raw History Blame Contribute Delete
3.21 kB

3DCodeBench extra ablation logs

Extra ablation / supplementary experiment logs for 3DCodeBench.

This folder is not the main leaderboard inference dump. Use the two log trees as follows:

Folder Role
3DCodeBench_ModelLogs/ Main eval logs of the reported models / agents (scripts, prompts, transcripts).
3DCodeBench_extra_ablation_logs/ (this folder) Extra ablations on top of that: curation, contamination, held-out categories, structural metrics, harness variants, interactive feedback, and track split. Archives include meshes and renders.

Each .tar.gz unpacks to a directory with the same stem. Virtualenvs and __pycache__ are omitted. Start at findings/ for writeups and tables.

Extra ablations

Path Extra ablation
findings/ Per-run writeups and tables.
e1_curation_ablation.tar.gz Curation ablation. Three settings: naive distill (A), agentic pipeline without human (B), released data with human verification (C). Result GLBs/PNGs vs references.
e1b_annotation (in small_notes.tar.gz) Annotator rubric and protocol notes.
e2_cad_pilot.tar.gz Same manufactured categories, Blender vs CAD (FreeCAD) generation.
e4_contamination.tar.gz Code-overlap analysis of model scripts against Infinigen sources.
e4c_fresh_categories.tar.gz Held-out / non-Infinigen categories, plus a copy of the original 212-class references, re-renders, and model outputs on both sets. Largest archive.
e5_structural.tar.gz Structural / topology metrics (floating parts, watertightness, fragmentation) and the meshes they were scored on. Scorer: findings/e5_structural/structural_integrity.py.
e5_render_pass.tar.gz Incomplete re-render pass. Stopped on purpose; see findings/e5_render_pass/ABANDONED.md.
e5_vote_agreement (in small_notes.tar.gz) Scoping notes for per-vote metric agreement.
e6_track_split (in small_notes.tar.gz) Text-to-3D vs image-to-3D split tables. Analysis only.
e8_harness.tar.gz Extra harness ablation on 50 categories: single-shot, a unified ReAct loop (same prompt/tools/turn cap for every backbone), and native Gemini CLI. Also a 20-turn rerun of the unified loop.
e9_interactive.tar.gz Multi-round generation with extra feedback (grounded vs model-only).
issue1_fixes (in small_notes.tar.gz) Dataset script fixes and diffs.
taxonomy (in small_notes.tar.gz) Category-group labels.
small_notes.tar.gz The small items above, plus the experiment plan.
backups_harness-storage.tar.gz Snapshot of harness working storage used while running the extra ablations.
SHA256SUMS Checksums of the archives.

e8 harness conditions (short)

Held fixed across backbones in the unified loop: system prompt, tools {write_file, run_blender, read_render}, 10-turn cap (20 in the follow-up), 600 s wall clock.

Condition Meaning
single_shot One-call generation
unified Neutral ReAct harness
unified_t20 Same harness, 20 turns
native_cli Vendor Gemini CLI

Restore

tar -xzf e8_harness.tar.gz
tar -xzf e4c_fresh_categories.tar.gz