section float64 5.1 5.4 | check stringlengths 35 108 | passed int64 1 1 | detail stringlengths 5 111 ⌀ |
|---|---|---|---|
5.1 | router_stability = 1.0 after the planted freeze | 1 | min post-freeze stability=1.0000 |
5.1 | crystallisation_step recovers the planted freeze (2000) | 1 | recovered=2200.0 |
5.1 | two INDEPENDENT routers look ~72% stable on raw Hamming but ~0 chance-corrected | 1 | raw=0.718 chance=0.722 adj=-0.013 |
5.1 | a router that never settles returns None (NEGATIVE CONTROL) | 1 | got None |
5.1 | planted segmentation-BEFORE-capability -> 'SEGMENTATION FIRST' verdict | 1 | SEGMENTATION FIRST -- fixed substrate; route |
5.1 | planted capability-BEFORE-segmentation -> 'CAPABILITY FIRST' (NEGATIVE CONTROL) | 1 | CAPABILITY FIRST -- circuits precede stabl |
5.1 | asynchrony_test detects planted late low-resource crystallisation | 1 | rho=0.88 p=0.0269 |
5.1 | tau1_window brackets the planted gap (1600, 4000) | 1 | window=(1800.0, 4200.0) |
5.1 | SYNCHRONOUS languages -> no asynchrony and NO principled tau1 (NEGATIVE CONTROL) | 1 | rho=0.00 p=1.000 window_valid=False |
5.1 | capability_inflection recovers a planted sigmoid inflection (3000) | 1 | got 3400.0 |
5.2 | pool_chunks(mean) is the chunk mean | 1 | null |
5.2 | span-restricted pooling selects only overlapping chunks | 1 | null |
5.2 | planted SHARED geometry -> high whitened alignment gain | 1 | gain=1.000 |
5.2 | planted UNRELATED languages -> ~zero gain (NEGATIVE CONTROL) | 1 | gain=-0.001 |
5.2 | retrieval P@1 after alignment: shared high, unrelated at null | 1 | shared=1.00 unrelated=0.03 |
5.2 | CKA/RSA (rotation-free) separate shared from unrelated | 1 | cka 1.00 vs 0.10, rsa 1.00 |
5.2 | injected anisotropy lowers IsoScore and inflates off-diagonal cosine | 1 | iso 0.061 vs 0.902; offcos 0.16 vs -0.00 |
5.2 | anisotropy_sensitive flag FIRES on anisotropic data | 1 | delta_gain=+0.762 |
5.2 | anisotropy_sensitive flag is QUIET on isotropic data (NEGATIVE CONTROL) | 1 | delta_gain=-0.000 |
5.2 | whitening recovers the true shared geometry under anisotropy | 1 | gain_white=1.000 |
5.2 | pairwise_table reports probe accuracy and a full anisotropy audit | 1 | probe_acc_white=0.30 chance=0.33 |
5.2 | width_sweep recovers a planted 'wider = more shared' trend | 1 | rho(N, sharing)=+1.00 |
5.2 | width_sweep also recovers the accompanying drop in separability | 1 | rho(N, separability)=-1.00 |
5.2 | sweep_verdict states the recovered trade-off in words | 1 | rho(N, separability) = -1.00; rho(N, sharing) = +1.00; wider chunks are MORE shared and LESS language-separated |
5.2 | compare_sources ranks planted sharing correctly (hnet > bpe > byte) | 1 | hnet:0.94 > bpe:0.42 > byte:-0.00 |
5.3 | chunk_index_at / aligned_sites / misaligned_offsets place patches correctly | 1 | null |
5.3 | patch_effect is signed so that POSITIVE = improved prediction | 1 | null |
5.3 | region restriction strips the trivial within-span self-reconstruction credit | 1 | whole-sequence=0.2774 vs after-patch-only=0.1585 |
5.3 | planted transferable content -> aligned patch improves prediction | 1 | effect=0.2342 CI=(0.19806601373115348, 0.27255066410129847) |
5.3 | content-MISMATCHED donor null is centred on ~zero | 1 | null=-0.0122 CI=(-0.029104634476045707, 0.004660881617882052) |
5.3 | aligned effect separates from the null (large d, p < 0.01) | 1 | d=2.13 p=0.0000 |
5.3 | null success rate sits at its 0.05 construction value | 1 | 0.050 |
5.3 | ALIGNED (boundary) patching beats MID-CHUNK patching | 1 | delta=0.0699 p=0.0290 d=0.40 |
5.3 | no-transfer model -> no significant effect (NEGATIVE CONTROL) | 1 | effect=0.00e+00 p=1.000 |
5.3 | compare_checkpoints detects planted B > A transferability | 1 | B-A=0.1966 d=1.52 p=0.0000 |
5.3 | identical checkpoints -> hypothesis NOT supported (NEGATIVE CONTROL) | 1 | B-A=0.00e+00 p=1.000 |
5.4 | UD treebanks present on disk (>= 20 languages) | 1 | 27 treebanks: am_att, ar_padt, cy_ccg, de_gsd, en_ewt, es_gsd... |
5.4 | load_ud reconstructs the surface string exactly (byte offsets are valid) | 1 | mean=0.9995, worst=fi_tdt 0.986 |
5.4 | SIGMORPHON 2022 loader keeps only SURFACE segmentations (offsets exist) | 1 | 40107 surface items, kept_frac=0.699 |
5.4 | MorphyNet inflectional loader yields valid morpheme offsets | 1 | 5000 items, surface_frac=0.718 |
5.4 | planted 85%-recall router scores high F1 above a rate-matched random baseline | 1 | F1=0.846 random=0.071 above_chance=0.834 |
5.4 | a RANDOM router at the same rate scores ~0 above chance (NEGATIVE CONTROL) | 1 | F1=0.051 above_chance=-0.027 |
5.4 | AUROC of a signal-carrying router score is high; a random score is ~0.5 | 1 | signal=0.907 random=0.556 |
5.4 | EMA acting only on the low-confidence band is distinguishable from a entropy-matched generic regulariser | 1 | KS_lowconf ema=0.283 generic=0.190 |
5.4 | a 'smoothing module' that is really just tempering is NOT distinguishable (NEGATIVE CONTROL) | 1 | KS_lowconf ema=0.190 generic=0.190 |
5.4 | level_typing calls the dense level morph-like and the sparse level word-like | 1 | L1=morph-like (m=-0.35), L2=word-like (m=+0.35) |
5.4 | unit_length_stats reports both byte and character widths (continuous-script languages need chars, not bytes) | 1 | {'mean_bytes': 12.64, 'median_bytes': 12.0, 'n_units': 42, 'mean_chars': 4.31, 'median_chars': 4.0} |
5.4 | alignment_predicts_downstream reports a weak/no relation as such (the MorphScore-70 caveat) | 1 | R2=0.072 p=0.253 |
5.4 | ...and still detects a genuinely strong relation (NEGATIVE CONTROL) | 1 | R2=0.931 p=0.0000 |
H-Net dynamic-chunking: experiment results
Every result table behind the study, including the ones that failed. 41 experiment directories;
each has a RESULTS.md (verdict + caveats) alongside its machine-readable CSV/JSON, and the
82.5M re-runs also carry an auto-rendered RESULTS_auto.md produced by the same report script
as the pilot tables.
Headline findings
| experiment | question | verdict |
|---|---|---|
exp21_tier1_pilot |
does a parity objective equalise chunk allocation? | each beta variant equalises the denominator it targets (A→B +62.1% cps; Apc→Bpc +51.9% cpc), attributable — EMA and generic-regularisation controls do not reproduce it |
exp26_crystallisation |
do boundaries stabilise before capability? | no — capability first, 36/36 language-seed pairs; 20/36 never crystallise by step 6000 |
exp27_chunk_content |
are learned chunks linguistic? | no — morpheme F1 at or below rate-matched chance; gold features add ≤0.001 AUROC over computational features in 9/9 languages |
exp25_chunk_patching |
are chunks functional computational units? | no — aligned-minus-misaligned at the construction rate (0/28 cells >2 SD), while killing the chunk pathway costs +0.34 nats/byte |
exp28_difficulty_gate |
does chunking track difficulty? | partial — boundary rate yes (+0.181 ± 0.057, 12/12 languages); chunk length no (wrong sign) |
exp28_compute_knob |
is there an inference-time compute knob? | no — one-sided; BPB is minimised at the training threshold and worse in both directions |
exp28_mainnet_ablation |
how much is the chunk network worth? | ~0.07 BPB, not the +0.51 a naive off-distribution ablation reports |
exp24_neural_chunkers |
how do released chunkers segment? | Bolmo-1B matches its distillation teacher at boundary F1 0.987 after 39.3B tokens |
exp29_gate1_neural → exp33_neural_panel_221 |
Gate 1 for released neural chunkers, on all 221 FLORES+ varieties | neural slope −0.035, inside the pre-registered [0, −0.08] window (exp29's −0.10 was a 68-variety sample effect); BLT's trained patcher −0.084 = XLM-R; Bolmo = its teacher to 2 dp; H-Net innermost ABSORBS (Gini 0.12 vs byte 0.21), Bolmo PROPAGATES (0.36) |
exp21_tier1_spec |
does the parity result survive 3.7x the parameters and the full 5 GB budget? | −54 % at 0.72 GB at ~0 cost, and condition C's lock-in reversion (+0.115) replicates to the decimal; but at 5 GB the router freeze does not bank the effect — the still-trainable encoder re-routes around the frozen router in 2/3 seeds (end Gini 0.10 / 0.18 / 0.28 vs 0.23 unregularised). The equalisation holds while the objective is on |
exp32_trained_controls |
what is the chunk-level network honestly worth at 82.5M? (P0, P1) | +0.068 ± 0.011 BPB at 0.72 GB (= the pilot's 0.07), +0.10 at 5 GB — 5× below the inference-time ablation; local byte-tower attention costs 0.023 BPB and leaves it unchanged |
Scale check at 82.5M (spec: A–E × 3 seeds × 11k steps; specfull: A/B × 3 seeds × 76k steps, 5 GB)
| experiment | verdict at 82.5M |
|---|---|
exp26_crystallisation_{spec,specfull,specfull_dense} |
capability first replicates for low/mid-resource languages (3 seeds at both budgets; th/my never crystallise in 76k steps); on the 500-step early grid at 5 GB the whitespace high-resource languages are segmentation-first (en/de/fi/zh +1.2–2.4k steps) — the two clocks are ordered by resource tier. Probe grid stated in every table |
exp27_chunk_content_{spec,specfull} |
null replicates at 0.72 GB; at 5 GB a word-/character-aligned router for whitespace scripts appears in 2 of 3 seeds (en word F1-above-chance +0.80 / +0.42 / +0.02) at identical BPB — reachable, not inevitable; Devanagari/Tamil stay sub-character and morphemes null in every seed |
exp25_trajectory (+ exp25_chunk_patching_{spec,specfull}) |
GATE F fails at every checkpoint from step 256 to 76k, 3 seeds at both budgets: contrast-minus-null within ±0.01 against a +0.2..0.4 chunk-pathway control; no sign consistent across seeds at any step |
exp28_scale |
the adaptive-compute negatives are not scale artifacts (3 seeds at 5 GB): main-net upper bound flat, knob one-sided, hard-vs-easy deletion +0.02 at 50 % against a +0.29 random-deletion cost |
exp30_chunk_geometry |
chunk-space cross-lingual alignability is weak but above null (+0.05 whitened gain, P@1 3x chance); language identity 100% decodable; chunk-width sweep flat. n_sent < d degeneracy of whitened CKA/Procrustes documented |
exp31_pabpe_controlled |
parity-aware vs classical BPE, same trainer/corpus/merges: in-objective Gini 0.074 -> 0.004, out-of-objective 0.371 -> 0.345 -- the equity is a property of the language list |
Reporting discipline
Pre-registered and applied throughout: every effect carries a null (shuffled, circular-shift,
Gaussian-donor, or content-mismatched donor as appropriate); spread is reported across
seeds, never pooled over items; boundary F1 always carries its rate-matched random
baseline; FLOP claims state the counting rule (block(T,d) = 24 d²T + 2dT², causal-halved)
and use measured chunk counts; both directions of every result were treated as publishable.
exp23_interp_validation validates the estimators against synthetic ground truth with a
negative control for each positive check (49/49 passing).
Limitations
Pilot: 16.2M/22.5M params, 6000 steps, 3 seeds, one hierarchy stage, final step only. The 82.5M
re-runs (*_spec, *_specfull) remove the single-checkpoint limitation (21 / 30 retained
checkpoints; seeds 1–2 at 5 GB add a 500-step early grid) and the 5 GB budget has three seeds
(exp32's 5 GB one-chunk control has one). Still one hierarchy stage throughout. Entropy is a byte 5-gram proxy; the model's own
predictive distribution was never logged.
Companion 22.5M pilot models, 82.5M spec models with retained trajectories (A–E × 3 seeds at 0.72 GB, A/B × 3 seeds at 5 GB, and the exp32 control runs) and probes.
- Downloads last month
- 80