gun unlearning data

Target concept: gun (29 firearm tags: gun, pistol, rifle, revolver, shotgun, handgun, sniper rifle, uzi submachine gun, mp5 submachine gun, bullet, ammunition, gun barrel, … — full list in manifest_curated.json).

What's on disk

Path	Size	Shards / rows	What it is
forget/	37 GB	90 tars	Forget set — images containing the gun concept. 447,370 samples.
retain/	92 GB	269 tars	Retain set — everything else, ~3× the forget set. 1,342,021 samples (of 1,357,557 requested; seed 42).
pairs/	433 MB	69 files	Counterfactual pair tables (parquet). pairs_train_clean.parquet = 1,823,249 rows.
pair_shards_seq/	52 GB	180 tars	Pairs pre-materialized into sequential tar shards (used by the seq_fast recipe).
concept_table/	30 GB	—	Per-concept lookup table from Stage 2/3.
manifest_curated.json	5.5 MB	—	Authoritative counts + shard list. Read this, not the config, for exact numbers.
results/	—	—	Training output dir referenced by the 1M juwels_gun.yaml preset.

Counts (from manifest_curated.json)

n_forget_total:  447,370
n_retain_total:  1,342,021   (retain_size_requested 1,357,557, seed 42)
n_shards:        22,021

The counterfactual pairs

Each row of pairs_train_clean.parquet is a matched pair:

• image_pos — image with the gun concept
• image_neg — near-duplicate image without it
• pair_id, target_concept, jaccard, similarity_score, match_weight, block_label, tags_pos/neg, context_pos/neg, split

Use pairs_train_clean.parquet, not pairs_train.parquet — ~19.4% of the unclean file's rows have invalid 10-digit keys.

How it's used in training (current sequential recipe, where we need KLD anchor)

Config: configs/unlearn/juwels_bf16_l2dist_kl_v2_seq_fast.yaml (experiment_name: bf16_l2dist_kl_v2_seq_fast_gun, max_train_steps: 7000).

Two streams, merged per batch

• Pair stream — pairs_train_clean.parquet rows, images pulled from the DataComp tars at pair_tar_dir: /p/data1/mmlaion/cabs/dataconcept_128m (pair_source: "tar", 48-shard cache, pair_shuffle: false). PairDataset → PairCollator → pixel_values_pos [B,C,H,W] + pixel_values_neg [B,C,H,W], prefixed pair_*.
• Retain stream — retain SFT examples. RetainDataCollator → input_ids, labels, images, image_sizes, modalities (a normal LLaVA next-token batch; no attention_mask — LLaVA rebuilds it after <image>-token expansion).

Batch mix: pair_batch_size: 2, retain_batch_size: 2 (1:1). Bad/corrupt images are replaced by zero tensors rather than crashing a rank.

Sequential phase schedule

Unlike the interleaved presets, seq_fast runs phase_mode: "sequential", phase_forget_frac: 0.5 — two blocks over the 7000 steps:

Block (steps)	Data used	Active losses
FORGET — first 50% (0–3500)	pair stream only	visual invariance (pos≈neg) + L2 param-locality
RETAIN — last 50% (3500–7000)	retain stream	retain-SFT CE + KL-locality (student≈teacher logits) + L2 param-locality

L2 locality is on throughout (it anchors params, needs no data); the retain forward is skipped entirely during the FORGET block.

Loss weights (unlearning_loss_adapter.py)

Term	Data it consumes	Weight
retain-SFT cross-entropy	retain input_ids/labels/images	1.0
pooled visual invariance (l2, squared-dist, mean-pool)	pair_pixel_values_pos/neg	1.0
KL locality	retain batch + frozen teacher	0.5
L2 parameter locality	trainable params vs start snapshot	0.1

Model / optimisation

• Base: llava-onevision-qwen2-7b-ov, model_max_length: 1024, image_aspect_ratio: "pad".
• Trainable: projector only. bf16, FSDP, 8 dataloader workers, no grad checkpointing.
• LR 1e-5 cosine, warmup 0.1, per_device_train_batch_size: 2, grad-accum 1, max_grad_norm 1.0.

Net effect on data

Forget pairs teach the model to stop distinguishing gun-present from gun-absent images (invariance); the retain block re-anchors normal behaviour via SFT + KL-to-teacher; L2 locality keeps weights close to the original throughout so retain performance doesn't drift.