jungter/tiny-aya-global-unlearning-artifacts

Tiny Aya Global unlearning artifacts (Lexsi hackathon, Track 2b)

Everything we created for the study "Gone, or Just Hidden? Auditing unlearning of a book and a skill in Tiny Aya Global" except the model weights (published separately) and anything containing MUSE book text. Write-up: docs/REPORT_AYA.md; code: https://github.com/Ungter/lexsi-unlearning/tree/aya.

Published models (all derived from CohereLabs/tiny-aya-global, CC-BY-NC-4.0):

• jungter/tiny-aya-global-muse-retain-only
• jungter/tiny-aya-global-unlearn-book-npo
• jungter/tiny-aya-global-unlearn-book-surgery-b50
• jungter/tiny-aya-global-unlearn-book-surgery-basedelta-b25
• jungter/tiny-aya-global-unlearn-book-undo
• jungter/tiny-aya-global-unlearn-algebra-undo

Folders

folder	what it is	how it was produced
attack_sets/paraphrase_math/, translation_math/ (fr/de/es), word_problems/	held-out algebra attack sets: paraphrased, translated and word-problem versions of the algebra TEST items, with manifests and rejected items	written by a model from the algebra test inputs, kept only if a deterministic sympy verifier agrees (attack_sets/_build/*.py)
attack_sets/_inputs/math_*.jsonl	the algebra forget/retain test inputs the attack sets were built from	attack_sets/_build/prep_inputs.py
math_splits/{arithmetic,algebra,precalculus,calculus}/level{1,2,3}/{train,dev,test}.jsonl	generated math problems with exact answers (forget = algebra, the others = retain topics)	src/lfm3u math generator (sympy, seed 0) with leakage removal; math_splits/precalculus/hendrycks/ = MATH precalculus items (Hendrycks et al., MIT)
results/audit/	TEST audit per model (direct, paraphrase, translation, jailbreak, tool, best-of-k, book verbatim/knowmem, utility)	scripts/10_audit.py
results/relearn/, abl/, weight_attacks/, local_abl/, jspace/	relearning (LoRA, all weights), abliteration, INT8/INT4/noise, direction ablation, J-lens readouts	scripts/12_relearn.py, 11_abliterate.py, 20_weight_attacks.py, 17_local_abl.py, 05_jspace.py
results/dev/, results/unlearn/, results/undo/, results/finetune/	dev-split sweep results and training logs (unlearning, UNDO distillation, target/retain fine-tunes)	scripts/08_dev_eval.py, 07_unlearn.py, 15_undo.py, 03_finetune.py
results/auditkit/	independent re-evaluation with the hackathon's AuditKIT (lm-eval + native vLLM)	scripts/22_auditkit_eval.py
results/surgery/	AI-agent weight-surgery traces: every step file (steps/*.py, outputs *.out), scores, edit lists, state, certificates, box logs	the guarded harness src/lfm3u/surgery/; see docs/SURGERY_AYA_*.md on GitHub
results/SCORECARD_AYA.md, AUDIT_AYA.md, scorecard.json	the scorecard (95% bootstrap CIs) and the audit tables	scripts/18_scorecard.py, scripts/13_audit_report.py
results/runs/*/	Vast.ai run logs (run.log and per-job logs)	scripts/vast/launch.py
results/vast/	Vast launch records (offer, GPU, price; no secrets)	scripts/vast/launch.py
results/verbatk/, results/VERBATK_AYA.md	verbatim book recall under 9 prompt attacks: per-item ROUGE-L F1 per attack and the recovery table (no generations, no prompts)	scripts/23_verbatim_attacks.py, scripts/verbatk_table.py
writeup/	copy of the write-up (REPORT_AYA.md) and the compiled paper (paper.pdf); the originals are on GitHub	docs/REPORT_AYA.md, paper/
figures/	report figures	scripts/21_report_figures.py
sweeps/	every sweep / attack spec file (hyperparameters per run)	input to scripts/vast/jobs/*.sh

File counts: attack_sets/_build 3, attack_sets/_inputs 5, attack_sets/paraphrase_math 8, attack_sets/translation_math 5, attack_sets/word_problems 5, figures 5, math_splits/algebra 9, math_splits/arithmetic 9, math_splits/calculus 9, math_splits/precalculus 12, results 5, results/abl 8, results/audit 17, results/auditkit 63, results/base 1, results/dev 82, results/finetune 9, results/jspace 51, results/local_abl 2, results/relearn 30, results/retain_A15 1, results/runs 77, results/surgery 180, results/target_A15 1, results/undo 2, results/unlearn 76, results/vast 95, results/verbatk 12, results/weight_attacks 15, sweeps 54, writeup 2.

What is NOT here, and why

MUSE-Books (Harry Potter) text is copyrighted and is not redistributed: no forget/retain passages, no verbmem or knowmem questions, no book attack sets, no model continuations of book passages. Excluded by rule: book attack sets and inputs, the verbatim-prompt paraphrases/translations, every generations*/, probe*/, final_probe/ folder, per-item result JSONL, weight deltas. Every remaining text file was scanned for any 12-word span of the full MUSE-Books data (forget and retain text, verbmem and knowmem rows). A JSON file with a hit is kept only as a copy in which every offending string (prompt or model continuation) is replaced by [removed: MUSE-Books text] and that passes a second scan; any other file with a hit is dropped. Counts: 42 x excluded by rule (book data, generations, probes, deltas, smoke tests); 13 x book-text strings redacted (sanitised copy kept); 12 x contains a 12-word span of the MUSE-Books data. Get MUSE from its authors (Shi et al., 2024, muse-bench/MUSE-Books).

Withheld models (not published): checkpoints that reproduce the book, relearn it, or are degenerate:

• the MUSE target aya-muse-target-A15 and every algebra-unlearned model derived from it (SAM-NPO, RMU c300, all-weight NPO, LoRA NPO, the algebra agent patch, the extra seeds): they reproduce the copyrighted MUSE-Books (Harry Potter) text verbatim (TEST ROUGE-L F1 ~0.89 for the target).
• book gradient ascent book_ga_lr1e-6_s100: 82% of the book returns after 10 LoRA relearning steps.
• localized (masked) NPO mnpo_book_lr1e-3 and RMU layer 18 book_rmu_l18_c20_a100: direct or relearned verbatim recall too high (LoRA relearning peaks 0.31 and 0.84 of the gap).
• SAM-NPO book book_npo_sam_r0.1_lr3e-5: degenerate by the project's rule (12% gibberish on the verbatim prompts).

Licenses

Data we created (attack sets, generated math splits, results, logs, traces, figures): CC-BY-4.0. The MATH precalculus items under math_splits/precalculus/hendrycks/: MIT (Hendrycks et al., 2021). Numbers in the results were measured on models derived from CohereLabs/tiny-aya-global (CC-BY-NC-4.0, Cohere Labs); the models themselves are CC-BY-NC-4.0 (non-commercial).