[etomoscow]

Title: Add K-FORGE: Fisher-guided initialization for preference-based unlearning

---

Summary

Adds K-FORGE (Kronecker-Fisher One-shot Rank-r Generalized Erasure), accepted at Findings of ICJNLP-AACL 2026, as a community method. K-FORGE is a training-free one-shot edit: it estimates empirical Kronecker-Fisher factors on forget and retain calibration batches, simultaneously diagonalizes the pairs, applies a retain-aware Wiener filter with a retain budget (lambda_tradeoff), truncates to rank r, and folds the negative edit into the checkpoint. The edited checkpoint can be used directly or as an initialization for NPO / SimNPO (their objectives remain unchanged). Follows up on #211.

Changes

• src/trainer/unlearn/kforge.py — new KFORGE trainer: calibration hooks (valid causal-LM loss positions, eval mode), factor damping (fixed / scale-relative adaptive), wiener_v2 (theorem-backed retain-aware edit) and legacy_v1 (original heuristic, kept for ablation) edit variants, optional spectrum logging, fp32 checkpoint serialization so small edits survive saving. Single-process calibration (multi-process raises a clear error).
• configs/trainer/KFORGE.yaml — defaults follow the paper configuration (wiener_v2, lambda_tradeoff=0.01, max_target_modules=4 as a memory guard); the per-run paper settings (rank=2, strength=0.45) are passed inline in community/methods/KFORGE/run.sh, following the convention of the other community methods.
• community/methods/KFORGE/ — README with method description, hyperparameters, results table, known boundaries; run.sh reproducing the two-stage flow (edit → downstream NPO; DOWNSTREAM_TRAINER=SimNPO supported).
• tests/test_kforge.py — CPU regression tests for the edit math: closed-form diagonal full-rank case (the paper's Theorem 1 reduction), λ=0 full-rank invariant (Δ = −α·W), causal-LM token masking, scale-relative factor damping.
• docs/links.md, README.md — method link and counts.

Results

Primary endpoint: Forget Q/A Probability (lower is better). TOFU forget10, Llama-3.2-1B-Instruct, n = 3 seeds, mean ± std; full tables in community/methods/KFORGE/README.md.

Steps	Method	Init	Utility ↑	Forget Prob. ↓
50	NPO	scratch	0.509 ± 0.004	0.093 ± 0.001
		K-FORGE (0.450)	0.565 ± 0.003	0.047 ± 0.003
100	NPO	scratch	0.569 ± 0.009	0.043 ± 0.002
		K-FORGE (0.450)	0.587 ± 0.002	0.029 ± 0.002
50	SimNPO	scratch	0.577 ± 0.006	0.662 ± 0.003
		K-FORGE (0.450)	0.572 ± 0.007	0.542 ± 0.002
100	SimNPO	scratch	0.582 ± 0.002	0.518 ± 0.010
		K-FORGE (0.450)	0.579 ± 0.003	0.413 ± 0.008

Compute-matched comparison with the edit's setup fully charged (FP32, evaluation excluded from timing): K-FORGE-initialized NPO reaches 0.046 forget probability vs 0.059 for scratch at lower total wall time (277.2 s vs 285.5 s) and fewer FLOPs (2.95e15 vs 3.78e15), while improving utility (0.577 vs 0.549).

Known boundaries (documented in the README and paper): direct relearning recovers the forgotten data; MUSE gains vary by metric; Qwen2.5-1.5B is unchanged at reported precision.

Testing

• python tests/test_kforge.py — CPU regression tests pass (edit math in float64).
• ruff check / ruff format --check clean on changed src files (repo defaults).
• End-to-end run.sh on a GPU host (Llama-3.2-1B-Instruct, TOFU forget10) — run output available on request; evaluator harness unchanged from upstream experiment=unlearn/tofu/default.

Checklist (per docs/contributing.md)

• Trainer implemented and registered (KFORGE)
• Exact commands documented (community/methods/KFORGE/run.sh)
• Method folder with README (hyperparameters, model-selection guidance, results)
• Leaderboard entry + HuggingFace unlearned checkpoints — will add after initial review (happy to discuss format)
• Paper link in docs/links.md

Reference

• Paper: K-FORGE: Fisher-Guided Initialization for Preference-Based LLM Unlearning, Findings of ICJNLP-AACL 2026.
• Paper code & experiments: https://github.com/etomoscow/kron_unlearn

---

Note: the upstream community/leaderboard.md entry and HF checkpoints are intentionally left for the review phase — happy to run whichever benchmark configuration the maintainers prefer for the leaderboard.

[molereddy]

Thanks for contributing!!

Added some nit comments and couple questions

1. can you confirm this was run e2e successfully? because eval uses 16 bit - for flash attn - and the train code converts model to 32 bit.
2. stage 2 loads model with bf16 so saving in 32 bit doesn't seem to make a difference. can you re-check how this works?

[etomoscow]

Fixed the nits in 1c750d0:

• ICJNLP → IJCNLP in docs/links.md and the method README.
• The trainer class now defaults to edit_variant=wiener_v2 / lambda_tradeoff=0.01, matching the trainer config.
• Removed configs/experiment/unlearn/tofu/kforge.yaml; run.sh now passes the paper settings as inline Hydra overrides, like the other community methods do.

About the dtype questions:

• edit_weight_dtype=float32 (the default) upcasts the model to fp32 before folding the edit, so the update runs as w_fp32 + δ and the artifact is saved in fp32. With preserve, the model stays bf16 and add_() runs in bf16. bf16 has an 8-bit mantissa, so a small α·δ can round to zero relative to w: the edit is lost at apply time, before any save.
• For a bf16-loading stage 2 you're right that one rounding happens either way (at load instead of at save). The default exists for the other consumers: the paper's standalone FP32 evaluations and the quantization audits read the saved artifact, and you can still override downstream dtype (torch_dtype) without the rounding baked into the artifact.
• e2e: confirmed. We produced the paper results end-to-end with this two-stage flow through the OpenUnlearning harness: stage 1 saves the fp32 artifact, stage 2 and the TOFU evaluation load it with the standard bf16/flash-attn model config (the cast happens at load), and the reported metrics come from that flow.

Happy to flip the default to preserve if you'd rather keep bf16 artifacts; otherwise I'll leave float32 as is.

[molereddy]

thank you for the clarification and fixes, please fix merge conflicts and we should be good to merge!

[etomoscow]

Merge conflicts resolved in 0bb804e