Datasets:

TokenBender
/

circuit-discovery

Tags:

License:

Dataset card Files Files and versions

xet

Community

circuit-discovery / README.md

TokenBender

Update circuit-shotting artifact index for issue 43

29d0b33 verified 3 days ago

preview code

raw

history blame contribute delete

13.1 kB

metadata

license: other
tags:
  - mechanistic-interpretability
  - circuit-tracing
  - arithmetic
  - qwen
  - sparse-circuits
pretty_name: Circuit Shotting Artifacts

Circuit Shotting Artifacts

This dataset stores large/generated artifacts externalized from Occupying-Mars/circuit-shotting.

The Git repo keeps source code, configs, lightweight reports, and current narrative documents. This HF dataset keeps generated datasets, raw result trees, pod artifact bundles, masks, dashboards, logs, adapters, and reference files that do not belong in normal Git history.

Artifact and Backup Locations

Issue 43 XCOMET Destructive MLP 95% Gate - 2026-05-18

Primary backup bundle: circuit-shotting/artifacts/xcomet_destructive_mlp_95/issue43_xcomet_destructive_mlp_95_20260518T145643Z
Mirror bundle: TokenBender/synth-data-en-pt-circuit (dataset) at issue43_xcomet_destructive_mlp_95_20260518T145643Z
Source repo/branch/commit: Occupying-Mars/circuit-shotting, tokenbender_experiments, 038f04f
Pod/run root: brave-hawk-1c, /root/runs/issue43_xcomet_destructive_mlp_95_038f04f
Model/task: tencent/HY-MT1.5-1.8B, EN->PT translation
Held-out eval: NTREX EN->PT, 1,012 rows
Judge: Unbabel/XCOMET-XXL via four local RTX 6000 Ada service workers
Method: restore #42 ranked 100x1%ile MLP dust damage rows; build cumulative flat percentile candidates topN in {10,15,20,30,40,50}; nuke each candidate inside the full HY-MT model; generate NTREX hypotheses; score each candidate with XCOMET-XXL.
Gate: target recovery <= 0.05, where recovery is candidate_XCOMET / no_mask_XCOMET.
Prior floor: #39 final_mix_top_3, 5,898 channels (~3%), XCOMET recovery 0.15275785974445857, destruction 0.8472421402555415.
No-mask XCOMET: 0.9281014436671856
Candidate XCOMET table:
- percentile_top_10: 19,660 channels, XCOMET 0.18256553325603309, recovery 0.19670859742946434, destruction 0.8032914025705357
- percentile_top_15: 29,491 channels, XCOMET 0.2529226868034234, recovery 0.27251620879292665, destruction 0.7274837912070733
- percentile_top_20: 39,321 channels, XCOMET 0.19338692443559438, recovery 0.2083683047312901, destruction 0.7916316952687099
- percentile_top_30: 58,982 channels, XCOMET 0.189112861187974, recovery 0.20376313653897224, destruction 0.7962368634610277
- percentile_top_40: 78,643 channels, XCOMET 0.2072743144836056, recovery 0.22333152900248426, destruction 0.7766684709975158
- percentile_top_50: 98,304 channels, XCOMET 0.2053742483847226, recovery 0.2212842677770569, destruction 0.7787157322229431
Outcome: no candidate reached the 95% destructive gate, and none beat the #39 3% destructive floor. The flat whole-model percentile ranking should be treated as a negative result; the next mining frontier should recurse inside the #39 mined region instead of scaling flat percentile unions.

Issue 42 Top-20 Destructive MLP 1%ile Proxy Search - 2026-05-18

Primary backup bundle: circuit-shotting/artifacts/percentile_proxy/issue42_percentile_proxy_20260518T110628Z
Mirror bundle: TokenBender/synth-data-en-pt-circuit (dataset) at issue42_percentile_proxy_20260518T110628Z
Source repo/branch/commit: Occupying-Mars/circuit-shotting, tokenbender_experiments, 8b46e31
Pod/run root: golden-orbit-d6, /root/runs/issue42_percentile_proxy_8b46e31
Model/task: tencent/HY-MT1.5-1.8B, EN->PT translation
Eval mode: teacher-forced HY-MT KL proxy on FLORES devtest EN->PT, 512 rows; no XCOMET judge run
Method: split the full HY-MT MLP down-proj input channel universe into 100 contiguous 1%ile chunks; nuke-score all chunks by KL; union the 20 most destructive chunks; compare all-MLP-zero, top-20 nuked, and native top-20 substrate-only
Final mix: percentile_top_20, 39,321 MLP channels, 0.1999969482421875 of the HY-MT MLP channel universe
Proxy summary: all-MLP-zero mean token KL 12.821661838961273; top-20 nuked mean token KL 13.415645710292592 (-0.046326590015529634 KL-gap recovery vs all-zero); native top-20 substrate-only mean token KL 10.892991802528064 (0.15042278143481724 KL-gap recovery vs all-zero)
Top 1%ile chunks: sector_0_chunk_0, sector_0_chunk_1, sector_0_chunk_2, sector_0_chunk_91, sector_0_chunk_4, sector_0_chunk_3, sector_0_chunk_23, sector_0_chunk_5, sector_0_chunk_20, sector_0_chunk_65, sector_0_chunk_56, sector_0_chunk_7, sector_0_chunk_58, sector_0_chunk_6, sector_0_chunk_62, sector_0_chunk_9, sector_0_chunk_57, sector_0_chunk_59, sector_0_chunk_18, sector_0_chunk_60
Outcome: the selected 20% dust is strongly destructive by proxy when removed, but weak as native substrate-only; it is more a destructive chokepoint candidate than a sufficient MLP MVC.

Issue 41 Mined MLP Substrate-Only Sufficiency - 2026-05-18

Primary backup bundle: circuit-shotting/artifacts/mined_substrate_only/issue41_mined_substrate_only_20260518T103722Z
Mirror bundle: TokenBender/synth-data-en-pt-circuit (dataset) at issue41_mined_substrate_only_20260518T103722Z
Source repo/branch/commit: Occupying-Mars/circuit-shotting, tokenbender_experiments, c4354b4
Pod/run root: gentle-raven-34, /root/runs/issue41_mined_substrate_only_c4354b4
Model/task: tencent/HY-MT1.5-1.8B, EN->PT translation
Held-out eval: NTREX EN->PT, 1,012 rows
Judge: Unbabel/XCOMET-XXL on one 96 GB RTX PRO 6000 Blackwell Workstation Edition GPU
Compared systems: native_mined_substrate_only and actdelta_rank256_substrate_only
Substrate: #39 final_mix_top_3, 5,898 MLP channels, about 3.0% of the HY-MT MLP channel universe
XCOMET summary: native substrate-only 0.2118839286819925 (0.22814712205183943 recovery); ActDelta rank256 substrate-only 0.22047376124696297 (0.23739626893520915 recovery)
Outcome: the 3% mined MLP substrate is highly destructive when removed but weak as a standalone MLP sufficiency circuit; ActDelta rank256 improves substrate-only recovery only slightly over native.

Issue 39 ActDeltaLoRA MLP MVC - 2026-05-18

Primary backup bundle: circuit-shotting/artifacts/actdelta_lora_mlp_mvc/issue39_actdelta_lora_mlp_mvc_20260518T011201Z
Mirror bundle: TokenBender/synth-data-en-pt-circuit (dataset) at issue39_actdelta_lora_mlp_mvc_20260518T011201Z
Source repo/branch/commit: Occupying-Mars/circuit-shotting, tokenbender_experiments, 56c9195
Pod/run root: noble-shark-57, /root/runs/issue39_actdelta_lora_mlp_mvc_a1baaaa
Model/task: tencent/HY-MT1.5-1.8B, EN->PT translation
Held-out eval: NTREX EN->PT, 1,012 rows
Judge: Unbabel/XCOMET-XXL via 7 local B200-backed service workers
Method: MLP-only 10% sector nuke scan -> 1% dust mining -> final mix -> ActDeltaLoRA activation-delta SVD rank ladder
Final mix: final_mix_top_3, 5,898 MLP channels, proxy recovery 0.10299166698225783
XCOMET summary: no-mask 0.9287162019683438; nuked final mix 0.14186869932268653; best ActDeltaLoRA rank 256 0.5883143888676001 (0.633470577578718 recovery)
Outcome: artifact-preserved full run; rank-constrained ActDeltaLoRA recovered substantial quality over the nuked substrate but did not reach the 80% or 90% recovery gates.

Issue 32 Preserved #28 EN->PT Artifact Replay - 2026-05-13

Primary replay bundle: circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141918Z
Source artifact repo: Occupying-Mars/hy-lora-conditions
Base model: tencent/HY-MT1.5-1.8B
Held-out eval: NTREX EN->PT, 1,012 rows
Judge: Unbabel/XCOMET-XXL
Validated scores: base fixed k160000 0.6278288431, r2 fixed k160000 0.8031356555, r8 fixed k160000 0.8145537329
#28 recorded comparison: base 0.6300551002, r2 0.7655800197, r8 0.8121489377
Note: low_rank_lens/k160_r4 is not present in Occupying-Mars/hy-lora-conditions; no r4 replay row is fabricated.

Primary artifact and backup surfaces:

surface	HF repo	contents
Consolidated circuit-discovery dataset	`TokenBender/circuit-discovery` (`dataset`)	Main home for circuit-shotting generated datasets, run bundles, masks, logs, summaries, manifests, and pod snapshots.
Circuit-discovery checkpoints	`TokenBender/circuit-discovery` (`model`)	Model checkpoints and adapter artifacts that are meant to be consumed as model repos.
EN->PT synthesis dataset	`TokenBender/synth-data-en-pt-circuit` (`dataset`)	Issue 30 DataDesigner/OpenRouter source synthesis and HY-MT teacher labels. Mirrored here under `circuit-shotting/data/synth-data-en-pt-circuit/issue30_deepseek_openrouter_hymt_20260512T222915Z`.
Source code and lightweight specs	`Occupying-Mars/circuit-shotting` (`GitHub`)	Code, configs, scripts, issue specs, and lightweight documentation.

Latest issue31 live backup:

path	contents
`circuit-shotting/artifacts/issue31/acdc_region_student_enpt/partial_blackwell_20260513T100813Z`	Partial live Blackwell snapshot for issue31 ACDC-only EN->PT region-student search, created `20260513T100813Z` UTC.

Upstream model weights are not duplicated in this artifact dataset. Restore tencent/HY-MT1.5-1.8B and Unbabel/XCOMET-XXL from their source Hugging Face model repos/cache, then restore this repo's run/data artifacts around them.

Layout

path	contents
`circuit-shotting/data/`	generated evaluation datasets and correctness sets
`circuit-shotting/results/`	local result directories, masks, dashboards, profiles, plots
`circuit-shotting/artifacts/`	pod logs, fetched run artifacts, tarballs, summaries, adapter weights
`circuit-shotting/references/`	reference files removed from Git
`circuit-shotting/manifests/`	checksum and size manifests for externalized files

Restore

Download the artifact subtree with:

hf download TokenBender/circuit-discovery --repo-type dataset --include "circuit-shotting/**" --local-dir hf_artifacts

To restore the old repo-relative layout, copy hf_artifacts/circuit-shotting/data, hf_artifacts/circuit-shotting/results, and hf_artifacts/circuit-shotting/artifacts into the repo root.

Manifest

The manifest at circuit-shotting/manifests/externalized_git_artifacts_2026-04-25.json records source paths, HF paths, byte sizes, and SHA-256 checksums.

Total payload in this manifest: 1642 files, 4768.22 MiB.

Skipped junk files are macOS .DS_Store and ._* resource-fork files.

Tencent EN->PT MVC Prep

Current bootstrap artifacts for the Tencent HY-MT EN->PT MVC prep run:

path	contents
`circuit-shotting/bootstrap/tencent_mvc_enpt/source_20260512T135647Z/circuit-shotting-working-tree.tar.gz`	source snapshot with the MVC split builder, service-backed XCOMET scorer, prep script, Lium bootstrap script, runner, and README updates
`circuit-shotting/bootstrap/tencent_mvc_enpt/runner_20260512T135559Z/tencent_mvc_runner-0.1.tar.gz`	pip-install startup runner for Lium pods whose logs work but SSH does not
`circuit-shotting/artifacts/tencent_mvc_enpt/partial_h100_20260512T140115Z/`	preserved interrupted H100 prep root: source splits, teacher hypotheses, ReLP masks, logs, manifest, tarball, checksum

The intended completed prep bundle path is circuit-shotting/artifacts/tencent_mvc_enpt/prep_v1/. It is still empty because no 2x RTX6000 executor was available after the failed port-unreachable pod attempts.

Circuit-Shotting Artifact and Backup Index

Issue #32 true #28 HY LoRA artifact replay: primary TokenBender/synth-data-en-pt-circuit/issue32_hy_lora_conditions_repro_20260513T141925Z; mirror TokenBender/circuit-discovery/circuit-shotting/artifacts/issue32/issue32_hy_lora_conditions_repro_20260513T141925Z; source artifacts Occupying-Mars/hy-lora-conditions; reproduces preserved k=160000 base XCOMET 0.6278288431 and strongest r8 XCOMET 0.8145537329; includes source artifact manifest/config/train summaries, preserved PEFT adapters and ReLP mask, NTREX dumps, XCOMET scored pools, summaries, logs, and checksums; excludes upstream HY-MT/XCOMET weights, HF caches, API keys, service tokens, and merged full checkpoints. Circuit-shotting run artifacts are kept in two places: TokenBender/synth-data-en-pt-circuit is the primary experiment-data dataset repo, and TokenBender/circuit-discovery mirrors specs, runbooks, and compact backups under circuit-shotting/ for recovery on fresh pods. Upstream model weights, Hugging Face caches, API keys, and service tokens are intentionally excluded from these backups.
Issue #32 final ReLP region-student apples-to-apples run: primary TokenBender/synth-data-en-pt-circuit/issue32_relp_region_student_apples_20260513T124959Z; mirror TokenBender/circuit-discovery/circuit-shotting/artifacts/issue32/issue32_relp_region_student_apples_20260513T124959Z; includes config/spec, runbook note, ReLP mask, FLORES teacher-hyp data, NTREX dumps, proxy metrics, XCOMET summaries, logs, SHA256SUMS, and the rank-128 region-student checkpoint.