Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('imagefolder', {}), NamedSplit('validation'): ('json', {}), NamedSplit('test'): ('text', {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

CircuitGen Private Artifacts Index

This repository contains the private CGDSL data plus the published artifacts for the netlist_to_layout_v1 v1 experiment lineage.

Latest repaired CGDSL overlay

For future preprocessing, training, and eval, prefer:

  • derived/cgdsl/20260307-081439-fix-open-netlistify-ams-overlay

This overlay copies forward the unchanged roots from:

  • derived/cgdsl/20260305-134840-fix-build-cgdsl-patched

and regenerates only these three roots after the 2026-03-07 repair pass:

  • open-schematics_hf
  • netlistify_schematic_images_hf
  • ams-net.github.io

Key files:

  • derived/cgdsl/20260307-081439-fix-open-netlistify-ams-overlay/README.md
  • derived/cgdsl/20260307-081439-fix-open-netlistify-ams-overlay/manifest.json
  • derived/cgdsl/20260307-081439-fix-open-netlistify-ams-overlay/REGENERATION_FIX_REPORT.md

Historical note:

  • the published netlist_to_layout_v1 12k-step run below still used derived/cgdsl/20260305-134840-fix-build-cgdsl-patched
  • the exact historical inputs for that run remain listed later in this README

Primary published directories

Reproduction / restart snapshot

Path: repro_snapshots/netlist_to_layout_v1_3train1eval_512valsubset_12000steps_fullmachine_posteval_2026-03-07

Use this directory when you want to restart or reproduce the run on another machine.

What it contains:

  • exact training/eval code used for the final published lineage,
  • environment snapshots,
  • exact cached splits/vocab files for restart without preprocess,
  • exact source-root manifest with hashes,
  • smoke-run QA metrics/configs,
  • machine-readable run recipe and timeline.

Key files:

  • repro_snapshots/netlist_to_layout_v1_3train1eval_512valsubset_12000steps_fullmachine_posteval_2026-03-07/README.md
  • repro_snapshots/netlist_to_layout_v1_3train1eval_512valsubset_12000steps_fullmachine_posteval_2026-03-07/run_spec/current_setup.json
  • repro_snapshots/netlist_to_layout_v1_3train1eval_512valsubset_12000steps_fullmachine_posteval_2026-03-07/restart_cache/cache_manifest.json
  • repro_snapshots/netlist_to_layout_v1_3train1eval_512valsubset_12000steps_fullmachine_posteval_2026-03-07/data_inputs/used_source_roots_manifest.json
  • repro_snapshots/netlist_to_layout_v1_3train1eval_512valsubset_12000steps_fullmachine_posteval_2026-03-07/file_manifest.json

Reproduction / restart snapshot for repaired-overlay rerun

Path: repro_snapshots/netlist_to_layout_v1_overlay_exact_3train1eval_914val_8000steps_fullmachine_posteval_2026-03-08

Use this directory when you want to restart or reproduce the cleaned repaired-overlay rerun on another machine.

What it contains:

  • exact training/eval code snapshot for the repaired-overlay rerun,
  • environment snapshots for the rerun machine,
  • exact repaired-overlay cache bundle for restart without preprocess,
  • exact repaired-overlay source-root manifest with hashes,
  • machine-readable run recipe and timeline for stage A, stage B, and post-train eval.

Key files:

  • repro_snapshots/netlist_to_layout_v1_overlay_exact_3train1eval_914val_8000steps_fullmachine_posteval_2026-03-08/README.md
  • repro_snapshots/netlist_to_layout_v1_overlay_exact_3train1eval_914val_8000steps_fullmachine_posteval_2026-03-08/run_spec/current_setup.json
  • repro_snapshots/netlist_to_layout_v1_overlay_exact_3train1eval_914val_8000steps_fullmachine_posteval_2026-03-08/restart_cache/cache_manifest.json
  • repro_snapshots/netlist_to_layout_v1_overlay_exact_3train1eval_914val_8000steps_fullmachine_posteval_2026-03-08/data_inputs/used_source_roots_manifest.json
  • repro_snapshots/netlist_to_layout_v1_overlay_exact_3train1eval_914val_8000steps_fullmachine_posteval_2026-03-08/file_manifest.json

Experiment / analysis package

Path: experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000

Use this directory when you want to analyze the run.

What it contains:

  • all published checkpoints,
  • all published training metrics,
  • all published eval metrics,
  • run logs from each phase,
  • config snapshots,
  • metadata summaries and checkpoint lineage,
  • per-shard debug outputs for the final 4-GPU full-machine eval.

Key files:

  • experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/README.md
  • experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/metrics/train_metrics_full_step0001_12000.jsonl
  • experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/metrics/eval_metrics_step4000_8000_12000.jsonl
  • experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/metrics/best_eval_by_spec.json
  • experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/metadata/run_summary.json
  • experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/metadata/checkpoint_lineage.json
  • experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/file_manifest.json

Repaired-overlay smoke package

Path: experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageA2000_step0001-002000

Use this directory when you want the 2k smoke run on the repaired overlay and exact-supervision cache.

What it contains:

  • the smoke checkpoint at step 2000,
  • the full stage-A training curve,
  • the async eval row and evaluator state for step 2000,
  • cache metadata and root-level audits for the repaired overlay cache.

Key files:

  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageA2000_step0001-002000/README.md
  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageA2000_step0001-002000/metrics/train_metrics_step0001_002000.jsonl
  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageA2000_step0001-002000/metrics/async_eval_metrics_step002000.jsonl
  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageA2000_step0001-002000/metrics/best_async_val.json
  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageA2000_step0001-002000/metadata/run_summary.json
  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageA2000_step0001-002000/file_manifest.json

Repaired-overlay main rerun package

Path: experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageB8000_step0001-008000

Use this directory when you want the main 8k cleaned-overlay rerun.

What it contains:

  • published checkpoints for steps 2000, 4000, 6000, 8000,
  • the full 8k training curve,
  • merged 4-GPU full-machine eval rows for steps 4000 and 8000,
  • logs from training, async eval, handoff, and full-machine eval,
  • per-shard debug outputs for the full-machine eval,
  • cache metadata and root-level audits for the repaired overlay cache.

Key files:

  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageB8000_step0001-008000/README.md
  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageB8000_step0001-008000/metrics/train_metrics_step0001_008000.jsonl
  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageB8000_step0001-008000/metrics/full_eval_metrics_step4000_8000.jsonl
  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageB8000_step0001-008000/metrics/best_full_eval.json
  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageB8000_step0001-008000/metadata/run_summary.json
  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageB8000_step0001-008000/file_manifest.json

Exact input data versions used by this run

A newer repaired overlay exists for future work at derived/cgdsl/20260307-081439-fix-open-netlistify-ams-overlay, but the list below remains the exact CGDSL source version used by the published 12k-step run.

The exact CGDSL source files used by this training/eval lineage are the versions currently committed under:

  • derived/cgdsl/20260305-134840-fix-build-cgdsl-patched/ltspice-spice-circuits_hf.jsonl
  • derived/cgdsl/20260305-134840-fix-build-cgdsl-patched/ltspice-spice-circuits-no-aug_hf.jsonl
  • derived/cgdsl/20260305-134840-fix-build-cgdsl-patched/open-schematics_hf.jsonl
  • derived/cgdsl/20260305-134840-fix-build-cgdsl-patched/netlistify_schematic_images_hf.jsonl
  • derived/cgdsl/20260305-134840-fix-build-cgdsl-patched/ams-net.github.io.jsonl
  • derived/cgdsl/20260305-134840-fix-build-cgdsl-patched/Circuitsense-6k_hf.jsonl
  • derived/cgdsl/20260305-134840-fix-build-cgdsl-patched/CircuitSense-100-TF_hf.jsonl
  • derived/cgdsl/20260305-134840-fix-build-cgdsl-patched/README.md

Exact input data versions used by the repaired-overlay reruns

The cleaned reruns published above use the repaired overlay:

  • derived/cgdsl/20260307-081439-fix-open-netlistify-ams-overlay/README.md
  • derived/cgdsl/20260307-081439-fix-open-netlistify-ams-overlay/REGENERATION_FIX_REPORT.md
  • derived/cgdsl/20260307-081439-fix-open-netlistify-ams-overlay/ltspice-spice-circuits_hf.jsonl
  • derived/cgdsl/20260307-081439-fix-open-netlistify-ams-overlay/open-schematics_hf.jsonl
  • derived/cgdsl/20260307-081439-fix-open-netlistify-ams-overlay/ams-net.github.io.jsonl

The exact cache metadata and retained-root audits for the reruns are published inside:

  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageA2000_step0001-002000/metadata/cache_meta.json
  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageA2000_step0001-002000/metadata/cache_root_audits.json
  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageB8000_step0001-008000/metadata/cache_meta.json
  • experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageB8000_step0001-008000/metadata/cache_root_audits.json

The hash manifest for those files is published at:

  • repro_snapshots/netlist_to_layout_v1_3train1eval_512valsubset_12000steps_fullmachine_posteval_2026-03-07/data_inputs/used_source_roots_manifest.json
  • repro_snapshots/netlist_to_layout_v1_3train1eval_512valsubset_12000steps_fullmachine_posteval_2026-03-07/data_inputs/used_source_roots.sha256
  • repro_snapshots/netlist_to_layout_v1_overlay_exact_3train1eval_914val_8000steps_fullmachine_posteval_2026-03-08/data_inputs/used_source_roots_manifest.json
  • repro_snapshots/netlist_to_layout_v1_overlay_exact_3train1eval_914val_8000steps_fullmachine_posteval_2026-03-08/data_inputs/used_source_roots.sha256

Which files are most useful for which job?

If you want to restart training/eval elsewhere

  • start with repro_snapshots/netlist_to_layout_v1_3train1eval_512valsubset_12000steps_fullmachine_posteval_2026-03-07/README.md
  • use repro_snapshots/netlist_to_layout_v1_3train1eval_512valsubset_12000steps_fullmachine_posteval_2026-03-07/code/netlist_to_layout_v1
  • restore repro_snapshots/netlist_to_layout_v1_3train1eval_512valsubset_12000steps_fullmachine_posteval_2026-03-07/restart_cache into a working netlist_to_layout_v1/cache/
  • resume from experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/checkpoints/step-012000.pt or an earlier checkpoint in experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/checkpoints
  • use repro_snapshots/netlist_to_layout_v1_3train1eval_512valsubset_12000steps_fullmachine_posteval_2026-03-07/run_spec/current_setup.json as the canonical restart recipe

If you want to restart the repaired-overlay rerun elsewhere

  • start with repro_snapshots/netlist_to_layout_v1_overlay_exact_3train1eval_914val_8000steps_fullmachine_posteval_2026-03-08/README.md
  • use repro_snapshots/netlist_to_layout_v1_overlay_exact_3train1eval_914val_8000steps_fullmachine_posteval_2026-03-08/code/netlist_to_layout_v1
  • restore repro_snapshots/netlist_to_layout_v1_overlay_exact_3train1eval_914val_8000steps_fullmachine_posteval_2026-03-08/restart_cache into a working netlist_to_layout_v1/cache_overlay_20260307_exact/
  • verify the exact overlay inputs with repro_snapshots/netlist_to_layout_v1_overlay_exact_3train1eval_914val_8000steps_fullmachine_posteval_2026-03-08/data_inputs/used_source_roots_manifest.json
  • resume from experiments/netlist_to_layout_v1/v1-overlay-exact-20260307-stageB8000_step0001-008000/checkpoints/step-008000.pt or another published repaired-overlay checkpoint
  • use repro_snapshots/netlist_to_layout_v1_overlay_exact_3train1eval_914val_8000steps_fullmachine_posteval_2026-03-08/run_spec/current_setup.json as the canonical restart recipe

If you want to analyze what happened in this run

  • start with experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/README.md
  • use experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/metadata/run_summary.json for the top-level summary
  • use experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/metrics/train_metrics_full_step0001_12000.jsonl and experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/metrics/eval_metrics_step4000_8000_12000.jsonl for curves
  • use experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/logs and experiments/netlist_to_layout_v1/v1-full-20260306-1900-async3g1e_step0001-012000/debug for debugging/performance analysis

Notes

  • The restart snapshot is the source of truth for reproduction/restart.
  • The experiment package is the source of truth for analysis/reporting.
  • The exact source JSONL versions used here are committed in this repo and fingerprinted here so the cache/checkpoint lineage is not dependent on this machine alone.

Historical note on max_steps

This published run lineage did not use a constant max_steps=12000 for the entire training run. Training began with max_steps=100000, following the original v1 plan. Near step ~11000, the run configuration was changed to max_steps=12000, and training then stopped at step 12000.

Interpretation guidance:

  • the final artifact state and some copied config files may show max_steps=12000
  • that value reflects the late-run edited state, not the schedule used for most of training
  • most of the run therefore occurred under the earlier 100000-step configuration
Downloads last month
165