README.md

metadata

pretty_name: Tiny Aya compression evaluation (numbers and ids)
license: other
license_name: per-source-notices
language:
  - en
  - hi
  - ar
tags:
  - evaluation
  - quantization
  - unlearning
  - safety
size_categories:
  - 10K<n<100K

Tiny Aya compression evaluation: numbers and ids only

Released 2026-11-14. Code and reproduction: https://github.com/Muditkumar123/tiny-aya-unlearning-4bit.

This dataset holds the per-question and per-prompt scores behind two studies of CohereLabs/tiny-aya-global under 4-bit compression:

• Study U: does TOFU unlearning survive qint4 quantization?
• Study S: does keeping a refusal circuit in bf16 change refusal behaviour, compared with plain qint4 and a same-size random choice?

It contains no prompts, gold answers, model answers or raw judge text. The prompts are rebuilt from their pinned public sources with the builder scripts in the code repository. Everything here is ids, labels and numbers.

Contents

Path	What	Rows
export/study_u/<model>/ (original, c0v2, c1, c2, c3)	TOFU forget/retain: id, author, ROUGE-L recall, gold/paraphrase/perturbed log-probs, truth ratio. Global-MMLU en/hi: id, subject, gold letter, letter log-probs, prediction, correct. Plus settings (manifest.json) and summaries.	400 + 400 + 500 + 200 per model
export/study_s/eval/<model>/ (original, plain-q4, random30-q4, protected-q4)	per prompt: id, language, HarmBench category or XSTest type, WildGuard labels as 0/1 (harmful request, refusal, harmful answer), KL to the original, estimated answer length	168 + 50 + 50 + 250 per model
export/study_s/mmlu/<model>/	Global-MMLU rows	500 + 200 per model
export/tables/	the result tables, as reviewed	—
ids/discovery_ids.json	the 32 HarmBench / Alpaca pairs used for circuit discovery: row ids and text sha256 only	32
ids/data_sha256.txt	sha256 of the 8 rebuilt prompt-set files, to check a rebuild	8
logs/	numeric run, training, merge and re-verification logs, with local paths shortened	—

Reproduce the tables (CPU)

From the code repository:

python Vedika/scripts/build_prompt_sets.py && python Vedika/scripts/build_tofu_sets.py   # optional: rebuild prompts, compare with ids/data_sha256.txt
python Vedika/release/reproduce.py <path to export/>

reproduce.py recomputes every table in export/tables/ from the rows and stops on any difference.

Sources and notices

This repository adds no licence to anyone else's material. The numbers refer to these sources, each under its own terms, as read from the dataset or model cards. That reading is not legal advice.

Source	Revision	Card licence	Used as	Citation (from its card)
walledai/HarmBench (standard)	fb6c2afd	MIT	ids hb_std_<row>	Mazeika et al. 2024, arXiv:2402.04249
CohereLabs/aya_redteaming	5a16fee0	Apache-2.0	ids aya_<lang>_<line>	Aakanksha et al. 2024, arXiv:2406.18682
Paul/XSTest	f600c994	CC-BY-4.0	ids and prompt type	Röttger et al. 2024, NAACL, aclanthology 2024.naacl-long.301, arXiv:2308.01263 (same title on arxiv.org); authors' code paul-rottger/xstest (paper and code as linked from the card at f600c994)
locuslab/TOFU	324592d8	MIT	ids and author index	Maini et al. 2024, arXiv:2401.06121
CohereLabs/Global-MMLU	0e619dbe	Apache-2.0	ids, subject, gold letter	Singh et al. 2024, arXiv:2412.03304
tatsu-lab/alpaca	dce01c9b	CC-BY-NC-4.0	row ids and hashes only (no text)	per its card
allenai/wildguard (judge)	—	Apache-2.0; gated, AI2 research-use terms	labels it produced	Han et al. 2024, arXiv:2406.18495
CohereLabs/tiny-aya-global and the derived models	af89d219	CC-BY-NC-4.0 + Cohere acceptable-use policy	the models measured	per its card

The scores describe outputs of Tiny Aya and its derivatives, so their non-commercial licence and acceptable-use policy apply to any use of those models. The team's own original numeric results and aggregate tables are offered under CC-BY-4.0, only to the extent of the rights the team can grant. This does not cover the third-party datasets, their ids or content, the toolkits or the model weights, which keep their own terms above, and it gives no permission for commercial use of the models.

Limits

• Judge labels. Refusal and harmfulness are WildGuard's labels, not human ratings, and they are not validated on Hindi or Arabic answers. A planned blind human check did not produce a validation. Two team members' annotation sheets for 140 answers (25 Hindi and 10 English prompts × 4 models) have identical labels, and whether they were rated independently is unconfirmed. They were used only in an exploratory comparison, documented in the code repository at Mudit/release/j3_public/J3_EXPLORATORY_RESULTS_20261114.public.md. There, 14 of 100 Hindi refusal labels and 10 of 100 Hindi harmful-answer labels differed from WildGuard's, all with the annotation yes where WildGuard said no; in English, 0 of 40 and 1 of 40 differed. The annotations themselves are not in this dataset. Arabic was not examined.
• Truncated answers. Many answers reach the 256-token limit (an estimate from re-tokenized text), so some were judged on truncated text.
• Small samples. 20 authors per TOFU set, 50 prompts per Aya language, and one random control in Study S.
• Selection and overlap. The C2 unlearned model was selected on the same TOFU rows. All TOFU rows scored are training rows, so for Study U, Global-MMLU is the only held-out utility measure. The Study S safety sets are not training data. HarmBench held-out also excludes the 32 prompts used for circuit discovery.
• Multiplicity. No correction. The headline cells are Study U's U2 forget retention and Study S's S1 HarmBench refusal; the other cells are supporting.
• Proxy metrics. ROUGE-L recall is word overlap, not factual recall. Size is parameter-memory accounting, not file or GPU memory.