"""Frozen interpretation protocol for the whitened + scale-up attractor libraries — pinned before the smoke test. This experiment runs the raw library's proven exemplar->label->validate pipeline (exp_01m0spvn1eew0sje5yh97cjpxe, its src/protocol.py constants are reproduced verbatim below) over every still-uninterpreted attractor pole: - the whitened GPT-2 library (exp_01m0tedzxhejkrq70x0pvcsay1, 895 axes) - the whitened Pythia/Qwen library (exp_01m0xrq2h1fdertpjmkd9x80cs, 2,168 axes) - the raw-metric Pythia/Qwen attractors (exp_01m0sy4fjhefv9tv0nwz9cx1nd, 466 axes, never labeled) 3,529 axes / 7,058 poles over 12 trained models. Random-init floors are not interpreted. Nothing downstream chooses a threshold at runtime. """ # ---- provenance ----------------------------------------------------------- CENSUS_EID = "exp_01m0rjapkveyqr5fytnhtwdzg8" # raw GPT-2 census (frame ref) RAW_INTERP_EID = "exp_01m0spvn1eew0sje5yh97cjpxe" # raw interpretation (pipeline) WGPT2_EID = "exp_01m0tedzxhejkrq70x0pvcsay1" # whitened GPT-2 census WNEW_EID = "exp_01m0xrq2h1fdertpjmkd9x80cs" # whitened Pythia/Qwen census RAWNEW_EID = "exp_01m0sy4fjhefv9tv0nwz9cx1nd" # raw scaling census # ---- models (registry copied from the whitened scaling census) ------------- # family: "gpt2" (HF GPT2, ln_2-input hook == census hook_resid_mid), # "neox" (parallel residual, block-input hook), "qwen3"/"qwen3_5" (SwiGLU, # post_attention_layernorm-input hook). bos = prepended sink token. ALL_MODELS = { "gpt2": dict(hf="openai-community/gpt2", family="gpt2", n_layers=12, d_model=768, d_mlp=3072, vocab=50257, bos=50256), "gpt2-medium": dict(hf="openai-community/gpt2-medium", family="gpt2", n_layers=24, d_model=1024, d_mlp=4096, vocab=50257, bos=50256), "gpt2-large": dict(hf="openai-community/gpt2-large", family="gpt2", n_layers=36, d_model=1280, d_mlp=5120, vocab=50257, bos=50256), "gpt2-xl": dict(hf="openai-community/gpt2-xl", family="gpt2", n_layers=48, d_model=1600, d_mlp=6400, vocab=50257, bos=50256), "pythia-160m": dict(hf="EleutherAI/pythia-160m", family="neox", n_layers=12, d_model=768, d_mlp=3072, vocab=50304, bos=0), "pythia-410m": dict(hf="EleutherAI/pythia-410m", family="neox", n_layers=24, d_model=1024, d_mlp=4096, vocab=50304, bos=0), "pythia-1.4b": dict(hf="EleutherAI/pythia-1.4b", family="neox", n_layers=24, d_model=2048, d_mlp=8192, vocab=50304, bos=0), "pythia-2.8b": dict(hf="EleutherAI/pythia-2.8b", family="neox", n_layers=32, d_model=2560, d_mlp=10240, vocab=50304, bos=0), "pythia-6.9b": dict(hf="EleutherAI/pythia-6.9b", family="neox", n_layers=32, d_model=4096, d_mlp=16384, vocab=50432, bos=0), "qwen3-0.6b": dict(hf="Qwen/Qwen3-0.6B", family="qwen3", n_layers=28, d_model=1024, d_mlp=3072, vocab=151936, bos=151643), "qwen3-8b": dict(hf="Qwen/Qwen3-8B", family="qwen3", n_layers=36, d_model=4096, d_mlp=12288, vocab=151936, bos=151643), "qwen3.6-27b": dict(hf="Qwen/Qwen3.6-27B", family="qwen3_5", n_layers=64, d_model=5120, d_mlp=17408, vocab=248320, bos=248044, dtype="bfloat16"), } DEFAULT_DTYPE = "float32" # 27B loads bf16 (spec dtype); rest fp32 # Pinned tag order — axis_gid / pole_id assignment depends on it. TAGS = ("gpt2", "gpt2-medium", "gpt2-large", "gpt2-xl", "pythia-160m", "pythia-410m", "pythia-1.4b", "pythia-2.8b", "pythia-6.9b", "qwen3-0.6b", "qwen3-8b", "qwen3.6-27b") # Axis sources per tag: GPT-2 tags carry only whitened axes (their raw axes # are the already-interpreted exp-4 library); new families carry whitened + # raw. Within a tag axes are ordered (layer, metric_order, cluster_id) with # metric_order whitened=0, raw=1. pole_id = 2*axis_gid + (0 pos | 1 neg). METRICS_BY_FAMILY = {"gpt2": ("whitened",), "neox": ("whitened", "raw"), "qwen3": ("whitened", "raw"), "qwen3_5": ("whitened", "raw")} # ---- corpus (census conventions, unchanged) -------------------------------- SEQ = 127 # tokens per sequence, BOS prepended at capture DATASET = "NeelNanda/pile-10k" NEW_SLICE_START_ROW = 363 # first pile row untouched by fit/held slices NEW_SLICE_TOKENS = 4_000_000 # per family, truncated to a multiple of SEQ # gpt2-family regression: the exemplar slice must reproduce the raw # interpretation run's slice exactly (its slice_meta sha). GPT2_SLICE_SHA = "bcf5eee2" # prefix pin; full sha asserted from staged meta CAPTURE_BATCH = {768: 64, 1024: 64, 1280: 32, 1600: 32, 2048: 32, 2560: 32, 4096: 16, 5120: 16} # harvest batch by width # ---- exemplar extraction (raw interpretation protocol, verbatim) ------------ TOP_ANCHORS = 128 # anchors retained per pole (cap-respecting walk) PER_SEQ_CAP = 2 # max anchors per 127-token sequence per pole PER_DOC_CAP = 4 # max anchors per pile-10k document per pole OUTLIER_SD = 4.0 # ||h|| > mean + 4 SD (new slice, per block) excluded WINDOW_RADIUS = 10 # +-10 tokens, clamped to the sequence N_LABEL_POOL = 60 # label-side anchors kept per pole (even rows only) N_LABEL_PROMPT = 20 # exemplars per label prompt, pool-sampled (seeded) N_DETECT = 10 # detection-true exemplars (odd rows only) N_NEG = 12 # non-activating negatives extracted per pole NEG_SEED = 43 # rejection-sampling seed (NEG_SEED + pole_id) NEG_MAX_QUANTILE = 0.5 # negatives drawn where signed proj < slice median NEG_PER_SEQ_CAP = 2 NEG_PER_DOC_CAP = 2 EXCL_TOP = 100 # pole positions a negative must stay away from EXCL_RADIUS = 10 MIN_LABEL = 8 # below these the pole is flagged insufficient MIN_DETECT = 6 # ---- per-model token profiles (native vocab; report display + sensitivity) -- PROFILE_MIN_OCC = 50 # ids with >= 50 occurrences on the new slice # ---- family clustering: HELD-slice word profiles (cross-tokenizer) ---------- # Profiles are the stored word-type profiles of the scale-up protocol # (rows 102-362 shared raw text, WordMap aggregation): exp-7 wpacks (whitened # new families), exp-7 wwordprof packs (whitened GPT-2), exp-5 arrays (raw new # families). Clustering uses the word types with >= WORD_MIN_OCC covered # occurrences in ALL FOUR tokenizer families (gpt2, neox, qwen3, qwen3_5). # All poles are clustered (signed profiles, pole = +-axis row); family naming # and kind categorization use validated labels only. WORD_MIN_OCC = 20 FAMILY_TAU = 0.70 # Spearman threshold (average linkage, cut 1 - tau) FAMILY_TAU_SENS = (0.65, 0.75) # stability band (plan: +-0.05) CROSS_MODEL_MIN = 3 # validated members from >= 3 of 12 models # (analogue of the raw run's >=3-of-4 rule) # ---- LLM stages (raw interpretation protocol, verbatim) --------------------- LLM_MODEL = "anthropic/claude-sonnet-5" # re-verified against the live catalog LLM_PROVIDER = "openrouter" REASONING_EFFORT = "low" LABEL_THREADS = 16 DETECT_THREADS = 16 DETECT_MAX_TOKENS = 2000 DETECT_RETRIES = 3 CONTROL_SEED = 42 # derangement of label -> pole assignments # Control derangement is WITHIN MODEL (labels never cross tokenizers or # content domains), never the same (model, layer) block — the within-GPT-2 # refinement of the raw run's block-forbidding derangement, needed because the # pooled library now spans three tokenizer families. ORDER_SEED = 42 # per-pole snippet shuffling (ORDER_SEED + pole_id) VALIDATION_PCTL = 95 # BA must beat this percentile of control BAs # Headline bar: pooled control q95 over all scored poles (raw-run convention). # Reported sensitivities: adjacent quantized bar, per-model bars, per-library # bars (whitened-gpt2 / whitened-new / raw-new). # ---- kind categorization ---------------------------------------------------- # Every validated pole label and every multi-member family name is classified # by the judge against the raw library's 15 mirrored sign-pair kinds (names + # descriptions extracted verbatim from exp-4's family_names/derived_stats into # data/raw_kinds.json at build time), plus a pinned coarse category set and an # explicit NEW option. Labels are classified in batches. KIND_BATCH = 20 KIND_COARSE = ("sentence_boundary", "function_word", "formulaic_continuation", "word_fragment", "punctuation", "indentation_whitespace", "noun_semantic", "other_new") # ---- reference checks (per model, before anchors are trusted) --------------- # Reference axis rule: among that model's whitened attractor-bearing blocks, # the block closest to n_layers/2 (ties -> lower); axis = largest share # (ties -> lowest cluster id). Recompute its held-slice profile in the # rebuilt frame; top-20 token overlap vs the stored census profile and # proj_top_mean relative error must pass. REF_TOKEN_OVERLAP = 0.75 REF_TOP_MEAN_RELERR = 0.02 REF_TOP_STATES = 800 # census pole_top_tokens convention REF_TOP_TOKENS = 20 # Smoke instrument reference: the raw census gpt2 L6 first attractor axis # (exp-4's known-interpretable kind-noun reference), rerun end-to-end. SMOKE_REF_BA = 0.70 SMOKE_BLOCKS = (6,) # gpt2 whitened block for the smoke SMOKE_TOKENS = 1_524_000 # 12,000 sequences # ---- constants referenced by the vendored families.py (unused paths) -------- RANDOM_INIT_SEED = 42 # random-init floors are NOT interpreted here STEER_MAX_NEW = 40 # steering is out of scope; kept for import STEER_TEMP = 0.9 STEER_TOP_P = 0.95 STEER_SEED = 42