sparse/protocol_sparse.py

"""Frozen constants for the sparse/semantic attractor probe — see
../PROTOCOL-SPARSE.md.  Pinned before any job runs.

Everything the frozen census pinned (token budget, SEQ, flow steps/seeds,
clustering threshold, count rule, colored floor, lift measurement) is
imported unchanged from `protocol.py` / `wprotocol.py`.  This module adds
only:

  1. the domain-dense corpus conditions (Q1) and their document rules,
  2. the depth-spanning block sets for the two censused models,
  3. the tail-concentration flow (Q2) — one changed ingredient, the on-set,
  4. the novelty rule against the models' existing Pile libraries,
  5. the pre-registered verdict rules.

Nothing here is chosen at runtime.
"""
import protocol as P  # noqa: F401  (re-exported constants are used verbatim)

# ==========================================================================
# provenance: where the frozen pipeline and the Pile reference come from

CENSUS_RAW_EID = "exp_01m0rjapkveyqr5fytnhtwdzg8"   # raw GPT-2 census
CENSUS_RAWNEW_EID = "exp_01m0sy4fjhefv9tv0nwz9cx1nd"  # raw Pythia/Qwen census
CENSUS_WGPT2_EID = "exp_01m0tedzxhejkrq70x0pvcsay1"  # whitened GPT-2 census
CENSUS_WNEW_EID = "exp_01m0xrq2h1fdertpjmkd9x80cs"   # whitened Pythia/Qwen
INTERP_EID = "exp_01m0zrgczdeemty0aep579zjnx"        # label/validate pipeline

# ==========================================================================
# 1. the four measures (corpus conditions)
#
# `pile` is the reference measure of every completed census; it is NOT rerun
# here — the stored per-block results are reused (identical slices, identical
# protocol).  The three new conditions are built with the census's own slice
# construction: documents in dataset order, fit until >= FIT_TOKENS, held
# from the first document the fit slice never touched, both truncated to a
# multiple of SEQ.  A document is a repo file / paper / conversation.

CONDITIONS = ("pile", "code", "math", "chat")
NEW_CONDITIONS = ("code", "math", "chat")

CORPORA = {
    "code": dict(
        hf="codeparrot/codeparrot-clean", config=None, split="train",
        field="content", kind="text",
        note="Python repository files; document = one file"),
    "math": dict(
        hf="math-ai/AutoMathText", config="arxiv-0.70-to-1.00", split="train",
        field="text", kind="text",
        note="arXiv papers with AutoMathText math score >= 0.70; "
             "document = one paper"),
    "chat": dict(
        hf="HuggingFaceH4/ultrachat_200k", config="default", split="train_sft",
        field="messages", kind="chat",
        note="UltraChat SFT dialogues rendered as a plain-text transcript; "
             "document = one conversation"),
}

# Plain-text chat rendering.  Deliberately NOT a model chat template: the
# measure must be the same *text* for both tokenizer families, so the role
# markers are ordinary words rather than special tokens.
CHAT_ROLE_NAMES = {"user": "User", "assistant": "Assistant",
                   "system": "System"}
CHAT_TURN_SEP = "\n\n"
CHAT_ROLE_SEP = ": "

# Streaming caps: how many documents we are willing to walk per split before
# declaring the corpus too short (guards a silent short slice).
MAX_DOCS_SCAN = 200_000
MIN_DOC_CHARS = 64          # skip degenerate/empty documents

# ==========================================================================
# 2. models and blocks
#
# The two most-baselined small models, plus their random-init floors.
# Blocks span depth at fixed fractions that reproduce each census's own
# representative blocks: round(L/6), round(L/3), round(L/2), round(5L/6).

MODELS = ("gpt2", "qwen3-0.6b")
FLOOR_OF = {"gpt2": "gpt2-randinit", "qwen3-0.6b": "qwen3-0.6b-randinit"}
TAGS = ("gpt2", "gpt2-randinit", "qwen3-0.6b", "qwen3-0.6b-randinit")
BASE_OF = {"gpt2": "gpt2", "gpt2-randinit": "gpt2",
           "qwen3-0.6b": "qwen3-0.6b", "qwen3-0.6b-randinit": "qwen3-0.6b"}

BLOCKS = {"gpt2": (2, 4, 6, 10), "qwen3-0.6b": (5, 9, 14, 23)}

# ==========================================================================
# 3. the tail-concentration flow (Q2)
#
# The frozen census flow ascends
#     Delta(v) = E[G(v; h) | h in top-q of a_v] - E[G(v; h)]
# with q = FLOW_Q = 0.1 re-drawn inside each 1024-state batch, i.e. it
# conditions on a ~10% event.  The tail flow ascends the SAME pointwise
# conductance under a rare-event conditioning: the on-set is the TAIL_M fit
# states with the largest signed projection under the CURRENT direction.
# Everything else — optimizer, learning rate, steps, batch, clamp, seeds,
# init directions, unit projection — is the frozen flow's, unchanged.
#
# TAIL_MODE selects how that on-set is found.  `exact` re-selects the corpus
# top-M every step (one chunked projection of the whole fit slice; a few
# seconds over the 200 steps at census scale), so the flow ascends the
# pre-registered functional itself.  `reservoir` is the cheaper persistent
# approximation the plan sketched; it only ever sees the states the flow drew
# (~40% of the corpus at this budget) and is kept as a robustness line.

TAIL_Q = 0.001              # rare-event quantile (0.1%), search and measure
TAIL_M = 399                # on-set size = max(30, round(TAIL_Q * N_fit))
                            #   pinned for the 399,923-state census slices
# Reporting-only diagnostic (not a decision rule): how many poles concentrate
# their top-M states into fewer than N distinct held sequences.  The library
# group is bimodal, so group medians are reported alongside these counts.
CONC_THRESHOLDS = (50, 100, 200)

TAIL_MODE = "exact"         # production selection
TAIL_MODE_SENS = "reservoir"  # robustness line, one block per model
TAIL_LIFT_Q = TAIL_Q        # held-out measurement quantile for the tail arm
TAIL_STEPS = P.FLOW_STEPS   # 200, frozen
TAIL_BATCH = P.FLOW_BATCH   # 1024 off-set draws per step, frozen
TAIL_LR = P.FLOW_LR         # 0.05, frozen
TAIL_CLAMP = P.FLOW_CLAMP   # 6.0, frozen
TAIL_SEED = P.FLOW_SEED     # 3, frozen (same batch stream as the raw flow)

# Concentration diagnostics measured for every accepted axis (and, as the
# comparison group, for the block's standard attractors).
CONC_TOPK = TAIL_M          # states in the measured tail
CONC_SEQ = P.SEQ            # a "moment" is a 127-token sequence

# ==========================================================================
# 4. novelty against the existing Pile libraries
#
# An attractor axis found here is NOVEL iff its centroid's largest absolute
# cosine against the union of the model's Pile attractor axes — raw census
# and whitened census, ALL blocks — is below JOIN_COS.  The bar is the
# clustering / identity threshold used by every census in this line of work.

NOVEL_COS = 0.9             # = P.CLUSTER_COS = wprotocol.JOIN_COS
NOVEL_FRAMES = ("raw", "whitened")   # reported in both; raw is primary
# Secondary (reported, not part of the rule): the join against ALL Pile
# cluster centroids, attractor or not, and against the same block only.

# ==========================================================================
# 5. exemplar extraction + labelling (interpretation pipeline, exp-8)

EXEMPLAR_TOKENS = 2_000_000     # per (family, condition), truncated to SEQ
PILE_EXEMPLAR_START_ROW = 363   # exp-8 NEW_SLICE_START_ROW (untouched rows)
LABEL_CATEGORIES = ("generic_syntactic", "domain_structural",
                    "semantic_topical")
CATEGORY_DESC = {
    "generic_syntactic":
        "domain-independent surface syntax: punctuation, sentence or line "
        "boundaries, function words, word fragments, whitespace, formulaic "
        "continuations — the kinds already catalogued in the Pile library.",
    "domain_structural":
        "structure specific to this domain's format: LaTeX environments or "
        "math mode, code block/indentation/bracket structure, identifiers as "
        "syntax, dialogue turn boundaries or role tags, markup.",
    "semantic_topical":
        "content rather than format: a topic, an entity or entity class, a "
        "mathematical concept, a subject-matter category.",
}

# ==========================================================================
# 6. pre-registered verdict rules  (../PROTOCOL-SPARSE.md section 6)
#
# Q1 (density).  Per condition c, over the 8 cells (2 models x 4 blocks):
#   n_novel(cell)       = novel attractor axes in the trained model, either
#                         census arm, passing that block/arm's own count rule
#   n_novel_floor(cell) = the same quantity for the random-init model of the
#                         same architecture under the same condition, joined
#                         against ITS OWN Pile library
#   supported iff  sum_cells n_novel >= 1  AND  one-sided Wilcoxon signed-rank
#   (trained > floor) over the 8 paired cells has p < ALPHA.
# Q1b (content).  Every novel pole that earns a detection-validated label is
#   categorised into LABEL_CATEGORIES by the judge; counts are reported with
#   exact binomial CIs.  "Semantics under a domain-dense measure" requires at
#   least one validated novel pole categorised semantic_topical.
# Q2 (sparsity in time).  Identical shape, with the tail census in place of
#   the condition census and the Pile measure throughout.

ALPHA = 0.05
N_PERM = 10_000             # permutation draws where a permutation null is used
PERM_SEED = 42
BOOT_N = 10_000
BOOT_SEED = 42
WILCOXON_ALT = "greater"    # trained > floor, one-sided

# Candidate-matched chance level for the novelty join: under a uniform random
# direction in d dimensions, cos^2 ~ Beta(1/2, (d-1)/2); the reported chance
# level is the expected maximum |cos| over the same number of library
# candidates (order statistic, Monte-Carlo with this seed).
CHANCE_SEED = 7
CHANCE_DRAWS = 20_000