ai-engram 0.6.0

Minimal, efficient covariance-based engram extraction for editing HuggingFace LLMs

ai-engram

Closed-form, covariance-based engram extraction for editing HuggingFace LLMs — forward-only, no gradient descent.

An engram is the slice of a layer's weights attributable to a target set of inputs. ai-engram isolates it analytically:

W_engram = W · Σ_target · pinv(Σ_total)

Σ_target and Σ_total are input covariances over the forget set and the reference set. Subtracting it — W ← W − α·W_engram — removes that knowledge while keeping the rest: fast, training-free unlearning / model editing.

• Closed-form — one pseudo-inverse per layer; no optimization loop, no labels.
• Forward-only — covariances via forward pre-hooks; no backprop.
• HF-native — Llama, Mistral, Qwen, Gemma, Phi … and GPT-2 (Conv1D) out of the box.
• Affine-correct — automatic bias absorption for bias-bearing layers.
• Tunable — per-layer edit scaling is pluggable: the paper's n/N (default), relative weight-norm, effective rank, or your own.

Statistics collection, closed-form extraction, and editing (apply / edit) are all here, and reproduce TOFU unlearning (see Validation).

Install

pip install ai-engram

Pulls torch, tqdm, and transformers — HF LLMs and GPT-2 work out of the box. Distribution name ai-engram; import name engram.

📖 Documentation: https://jeakwon.github.io/ai-engram/

Quickstart

Any nn.Linear (or GPT-2 Conv1D) model:

import torch
from engram import EngramEditor, EditorConfig

editor = EngramEditor(model, EditorConfig())

target = editor.collect_statistics(forget_loader)   # Statistics: mean covariance + counts
total  = editor.collect_statistics(total_loader)    # over the reference set

edited = editor.edit(target, total, alpha=1.0)      # compute the engram and subtract it
# or split it: engram = editor.compute_engram_weights(target, total); editor.apply(engram, alpha=0.6)

HuggingFace LLM (answer-token masked)

from engram import EngramEditor, EditorConfig

editor = EngramEditor(model, EditorConfig())

batch_fn = lambda b: {"input_ids": b["input_ids"], "attention_mask": b["attention_mask"]}
mask_fn  = lambda b: b["labels"] != -100           # covariance over answer tokens only

g_forget = editor.collect_statistics(forget_loader, batch_fn=batch_fn, mask_fn=mask_fn)
g_total  = editor.collect_statistics(total_loader,  batch_fn=batch_fn, mask_fn=mask_fn)

edited = editor.edit(g_forget, g_total, alpha=0.6)   # default scaling = the paper's n/N
# selective per-layer strength, e.g. by relative weight-norm:
#   from engram import weight_norm, compose, count_ratio
#   edited = editor.edit(g_forget, g_total, alpha=1.0, scale=compose(count_ratio(1.0), weight_norm(1.0)))

Restrict the edit to specific modules with target_modules — the same convention as LoRA/PEFT (["down_proj"] by name suffix, or a regex string), plus layers_to_transform for decoder-layer indices. See the Quickstart guide for details.

Mixture-of-experts. Answer-token masking reaches the experts automatically on transformers <5; on transformers ≥5 (fused experts) opt in to the detachable engram.moe adapter — EngramEditor(model, adapters=[FusedExpertAdapter()]) — covering ~35 fused MoE architectures (Mixtral, Qwen2/3/3.5-MoE, DeepSeek-V3, GLM4-MoE, MiniMax, Mistral4, OLMoE, Phi-MoE, …).

How it works

step	what	cost
1. collect	forward pre-hooks accumulate the mean xᵀx + sample count per layer	one forward pass, no backward
2. compute	projection P = W · C_target · pinv(C_total)	one pseudo-inverse per layer
3. apply	W ← W − α · f_l · P with a pluggable per-layer scaling f_l	a single subtraction

Efficient by construction — forward-only hooks, magnitude-bounded running-mean accumulation, CPU/GPU split (covariances on storage_device), a float32 solve cast back to the model dtype, and answer-token masking. The per-layer edit weighting f_l is pluggable (count_ratio default = the paper's n/N, weight_norm, effective_rank, …). Handles nn.Linear, GPT-2 Conv1D (a transposed linear), and masked variants; full details in the Guide.

Configuration (EditorConfig)

field	default	purpose
storage_device	model's device	where covariances are held; set "cpu" if the D×D matrices don't fit in VRAM (large models)
absorb_bias	True	absorb bias into the edit for bias-bearing layers

Validation

On TOFU forget10 with tofu_Llama-3.2-1B-Instruct, the engram extraction reproduces the paper's 14-metric Overall within ~0.01:

condition	ai-engram	paper
gold (retain90)	0.998	0.998
plain (α=0.6)	0.706	0.698
adaptive-norm (α=1.0, p=1)	0.817	0.818

Answer-token NLL confirms strong, selective forgetting — the forget set's NLL jumps ~16× while retain is preserved, and adaptive-norm beats plain on both axes. Runnable end-to-end in tests/ and examples/; see the TOFU page.

API

• collect_statistics(loader, target_modules=None, batch_fn=None, mask_fn=None, layers_to_transform=None) -> Statistics
• compute_engram_weights(target, total) -> EngramResult · apply(engram, *, alpha=1.0, scale=count_ratio(1.0)) -> Module · edit(target, total, *, alpha, scale)
• scaling functions: count_ratio · weight_norm · effective_rank · uniform · compose
• merge_statistics(*stats) · save_statistics(stats, path) · load_statistics(path)

Full reference (auto-generated from docstrings): API docs.

License

MIT © Jeakwon Kim