daniil2107/sajaniemi-prepare-cache

sajaniemi-role-extractor prepare-cache bundles (Gemma-3-12B, Llama-3.1-8B, Qwen3.5-9B)

Each <slug>.tar = configs + complete outputs/<slug>/cache/ tree + validation/timing/provenance (see MANIFEST.txt inside; sha256 alongside). Extract under the repo root of Pendigard/sajaniemi-role-extractor (master @ 8307646) next to data/, then run run-metrics with the bundled config.

prepare-cache campaign for the coauthor (Célian) — status written 2026-09-14 ~23:45 UTC

What runs

Pipeline: Pendigard/sajaniemi-role-extractor master @ 8307646 (repo/), pinned venv (venv/, uv.lock: torch 2.12.1+cu130, transformers 5.9.0, TransformerLens 3.3.0, SAELens 6.44.4). Dataset extracted under repo/data/sajaniemi_variable_dataset/. Per model, one chain per QoS lane (-g guaranteed-nils, -n normal+requeue): smoke -> code_1K -> code_1K_obfuscated_no_comment. Twin jobs never run the same (model, corpus) at once (locks/ protocol; done-markers in locks/*.done).

slug	model / SAE	layers	L0 (nominal)	shards	host mem
gemma-3-12b-pt_gs2-res-65k-medium	google/gemma-3-12b-pt + Gemma Scope 2 65k medium (Ben's config)	12/24/31/41 of 48	52/60/60/60	5	220G
llama-3.1-8b_llamascope-lxr-32x	meta-llama/Llama-3.1-8B + Llama Scope 131k (TopK-50 -> JumpReLU)	8/16/21/27 of 32	50	4	120G
qwen3.5-9b-base_qwenscope-w64k-l100	Qwen/Qwen3.5-9B-Base + Qwen Scope 64k TopK-100 (TransformerBridge)	8/16/21/27 of 32	100	4	72G
gemma-3-12b-pt_gs2-resall-262k-big (extra, after Gemma medium)	Gemma Scope 2 res-all 262k big	12/24/31/41	105/120/120/120	6	220G

Configs: configs/<slug>/{code_1K,code_1K_obfuscated_no_comment,smoke}.json (generator make_configs.py; Gemma SAEs served from local bf16 exports in saes/ because SAELens' gemma_3 loader needs network even offline; cache identity unchanged). Job ids: logs/jobs-g.tsv, logs/jobs-n.tsv (Gemma resubmitted with 220G: see the last lines).

Smoke results (5-file corpus incl. the longest code_1K file, 4.7k tokens)

• Qwen 31731: prepare 38 s, mean_l0 100.0, GPU ~16 GB, relocation passed.
• Llama 31728: prepare 4:47, mean_l0 79 (train) / 441 (test, 2 short files) — no-BOS sink effect, see caveats; RSS 72.1 GiB; relocation passed.
• Gemma medium 31744: prepare 1:53, mean_l0 59/53, GPU 34 GB, RSS 167 GB; relocation passed. (31725 OOM-killed at 100G.)

Caveats to pass on to the coauthor / Ben

1. Tokenization has no BOS (pipeline policy). The first token of every file becomes the attention sink; SAEs fire tens of thousands of features there (Llama: ~70k, max activation ~1360; Gemma max act ~13k). Exclude token_index 0 in run-metrics or prepend BOS (changes alignment offsets). Affects their 1B/4B runs identically. MEASURED on Llama code_1K train chunk 0, layer 8 (300,422 tokens, 246 files): token_index 0 has 70,814 active features on average (max activation 1,368) vs 29.7 for every other token (max activation 41); 66% of all stored (token, feature) pairs are these sink rows. The per-layer mean_l0 89-107 in *-validation.json is ~30 + sink; excluding token 0 it is ~30.
2. SAELens' llama_scope loader rescales the JumpReLU threshold by the dataset norm but ignores OpenMOSS' unit-decoder-norm transform -> effective L0 ~27-30 instead of 50 (review measurement). Observed per-layer mean_l0 is in *-validation.json.
3. Gemma 65k medium is L0 52-60 (Ben's choice); the closest Gemma option to "L0 ~100" is res-all 262k big (105-120) = the extra chain.
4. Qwen residuals come from TransformerBridge (bf16, eager attention, torch gated-delta-rule fallback), rev 68c46c4b…; SAE rev 82761c1c….
5. num_forward_layers raised from 1 to 4 (Llama 2): not part of cache identity, saves 4 forward passes per file.

Morning checklist

squeue --me -o '%.8i %.30j %.4t %.16q %.10M %R' | grep srx-
ls ~/sajaniemi_prepare_cache/locks/            # *.done = finished (model, corpus)
tail -n 30 ~/sajaniemi_prepare_cache/logs/srx-*-code_1K*_*.log
tail -f ~/sajaniemi_prepare_cache/repo/workspace/runs/<slug>/code_1K-<job>-prepare.log
ls ~/.titan-notices/                              # GPU over-budget warnings would land here
quota -s

Package for handoff (per model, once locks/-code_1K.done and -obf.done exist): ~/sajaniemi_prepare_cache/package_handoff.sh -> handoff/.tar (+ .sha256, MANIFEST.txt, STATUS.txt)

Completion record (2026-09-15 04:10 UTC)

All six priority deliverables are complete and validated (prepare + inspect; smoke-level relocation proof per model):

slug	code_1K	obfuscated_no_comment	bundle
gemma-3-12b-pt_gs2-res-65k-medium	job 31745 (49 min), L0 49-59	job 31746 (45 min), L0 46-56	handoff/gemma-3-12b-pt_gs2-res-65k-medium.tar 9.2 GB
llama-3.1-8b_llamascope-lxr-32x	job 31729 (54 min), L0 89-107 (~30 + sink)	job 31730 (62 min), L0 90-109	handoff/llama-3.1-8b_llamascope-lxr-32x.tar 10.4 GB
qwen3.5-9b-base_qwenscope-w64k-l100	job 31732 (66 min), L0 100.0	job 31733 (90 min), L0 100.0	handoff/qwen3.5-9b-base_qwenscope-w64k-l100.tar 16.8 GB

Each tar: configs/.json, runs/*-validation.json + *-timing.txt (+ smoke relocation verification), MANIFEST.txt, STATUS.txt, EXPORT_PROVENANCE.txt, and outputs//cache/{sae,annotations,span_feature}/ (complete). sha256 alongside. Extra chain (Gemma 262k big, L0 105-120) started its smoke at 04:07 UTC; results will be appended below if they finish.

Extra chain DONE 2026-09-15: gemma-3-12b-pt_gs2-resall-262k-big code_1K job 31748 (79 min, L0 88-116) and obfuscated job 31749 (79 min); bundle handoff/gemma-3-12b-pt_gs2-resall-262k-big.tar. Campaign complete: 8/8 full runs + 4/4 smokes, all on the normal lane, no preemptions.