stream_id stringlengths 20 30 | ct float32 3 108 | T float32 4 36 | C float32 12 3.89k | seed int16 0 19 | arm stringclasses 2
values | contingent bool 2
classes | shift_frac float32 0.2 0.8 ⌀ | n_cs_on int16 476 544 | n_us int16 480 548 | bayes_acq_full int16 0 26 | bayes_acq_truncated int16 0 228 | types listlengths 1.02k 1.02k | dt listlengths 1.02k 1.02k | target_logwait listlengths 1.02k 1.02k |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
ct3_T4_s0_contingent | 3 | 4 | 12 | 0 | contingent | true | null | 512 | 512 | 3 | 4 | [
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2... | [
1.7297375202178955,
5.554611682891846,
3.5603442192077637,
0.7039355039596558,
6.668478012084961,
3.6480205059051514,
0.9267928004264832,
0.2718348503112793,
5.99648904800415,
1.0165115594863892,
12.722740173339844,
1.9416837692260742,
0.10014637559652328,
10.396041870117188,
24.22676658... | [
1.7146285772323608,
1.4502732753753662,
-0.35106852650642395,
2.3337442874908447,
1.294184684753418,
0.1811772733926773,
-1.302560567855835,
1.9477657079696655,
0.016376769170165062,
2.685424327850342,
0.6635555028915405,
2.3510122299194336,
2.3414251804351807,
3.375814914703369,
1.61369... |
ct3_T4_s0_shuffle_control | 3 | 4 | 12 | 0 | shuffle_control | false | 0.426967 | 529 | 495 | 0 | 0 | [
0,
2,
0,
2,
2,
2,
2,
0,
2,
2,
0,
0,
0,
2,
0,
0,
2,
0,
0,
0,
0,
2,
2,
0,
2,
0,
0,
2,
0,
0,
0,
0,
2,
0,
0,
2,
2,
2,
0,
2,
0,
0,
2,
0,
2,
0,
2,
0,
2,
0,
0,
2,
2,
0,
0,
0,
2,
0,
2,
2,
2,
0,
0,
2... | [
3.9374797344207764,
3.346869707107544,
2.971881151199341,
1.2923986911773682,
10.316498756408691,
1.1986275911331177,
7.013000965118408,
12.047550201416016,
2.6168735027313232,
10.496188163757324,
0.23512403666973114,
10.878013610839844,
7.922082424163818,
10.212886810302734,
8.878533363... | [
1.2080254554748535,
1.4502732753753662,
0.2564999759197235,
2.3337442874908447,
0.1811772733926773,
1.9477657079696655,
2.685424327850342,
0.9619803428649902,
2.3510122299194336,
3.375814914703369,
3.367743492126465,
2.8978419303894043,
2.323650360107422,
3.4268221855163574,
3.0864932537... |
ct3_T4_s1_contingent | 3 | 4 | 12 | 1 | contingent | true | null | 512 | 512 | 2 | 3 | [
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2,
0,
2... | [
3.7140228748321533,
1.179775357246399,
4.734681606292725,
2.6762855052948,
40.21227264404297,
1.9116195440292358,
3.848917007446289,
1.400729775428772,
18.77414321899414,
1.5894646644592285,
0.9204654693603516,
2.8733956813812256,
3.9105348587036133,
3.716132164001465,
3.7914679050445557... | [
0.16532401740550995,
2.0029609203338623,
0.9844298362731934,
3.740615129470825,
0.6479508280754089,
1.6581608057022095,
0.33699339628219604,
3.013749361038208,
0.4633972644805908,
1.3333842754364014,
1.0554945468902588,
2.031651020050049,
1.312683343887329,
1.4513527154922485,
-0.7393637... |
ct3_T4_s1_shuffle_control | 3 | 4 | 12 | 1 | shuffle_control | false | 0.758512 | 529 | 495 | 0 | 0 | [
0,
0,
2,
0,
0,
2,
0,
0,
0,
2,
2,
0,
0,
2,
0,
2,
2,
0,
2,
2,
0,
0,
0,
0,
2,
2,
0,
2,
2,
0,
2,
0,
0,
0,
0,
0,
0,
0,
2,
0,
0,
2,
0,
2,
2,
2,
0,
2,
2,
0,
0,
0,
2,
0,
0,
2,
0,
2,
0,
0,
2,
2,
2,
0... | [
0.7699669003486633,
0.40881580114364624,
3.715015411376953,
0.4632980227470398,
6.9109978675842285,
0.036671001464128494,
9.382258415222168,
20.222986221313477,
12.492865562438965,
0.025781508535146713,
5.2496466636657715,
8.188776016235352,
11.63833999633789,
0.5364921689033508,
3.13445... | [
1.4167826175689697,
1.3123828172683716,
2.0029609203338623,
1.938406229019165,
-3.3057689666748047,
3.740615129470825,
3.4886474609375,
2.527219295501709,
-3.658097743988037,
1.6581608057022095,
3.013749361038208,
2.499370813369751,
-0.6227033138275146,
1.3333842754364014,
-0.41641902923... |
ct3_T4_s2_contingent | 3 | 4 | 12 | 2 | contingent | true | null | 512 | 512 | 5 | 75 | [0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2(...TRUNCATED) | [2.955296516418457,1.4947632551193237,14.030762672424316,1.2344173192977905,17.603483200073242,7.358(...TRUNCATED) | [0.4019678235054016,2.725574493408203,0.21059904992580414,3.2173585891723633,1.9958720207214355,2.45(...TRUNCATED) |
ct3_T4_s2_shuffle_control | 3 | 4 | 12 | 2 | shuffle_control | false | 0.691918 | 517 | 507 | 0 | 0 | [2,0,2,0,0,0,0,2,2,0,0,2,2,2,0,0,2,2,0,2,2,0,0,0,2,0,0,2,0,2,2,0,0,2,2,0,2,0,2,2,2,0,2,2,0,2,0,0,2,2(...TRUNCATED) | [4.45005989074707,13.647181510925293,1.6179983615875244,2.215902090072632,1.0985252857208252,10.0016(...TRUNCATED) | [2.725574493408203,0.4811898469924927,3.2173585891723633,3.1243979930877686,3.0748980045318604,2.454(...TRUNCATED) |
ct3_T4_s3_contingent | 3 | 4 | 12 | 3 | contingent | true | null | 512 | 512 | 3 | 3 | [0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2(...TRUNCATED) | [13.279826164245605,0.09898413717746735,11.891237258911133,5.679540157318115,3.1404149532318115,1.81(...TRUNCATED) | [-2.312795639038086,2.866237163543701,1.736870288848877,1.6009212732315063,0.597287118434906,2.49734(...TRUNCATED) |
ct3_T4_s3_shuffle_control | 3 | 4 | 12 | 3 | shuffle_control | false | 0.694625 | 514 | 510 | 0 | 0 | [0,0,2,0,0,2,2,0,0,2,2,2,0,2,0,0,2,0,2,2,2,2,0,2,0,0,2,0,2,0,2,0,0,0,2,0,2,0,2,0,2,0,2,0,2,0,0,2,0,0(...TRUNCATED) | [0.2551209330558777,8.59135913848877,4.53233003616333,7.636680603027344,5.584237098693848,4.34985923(...TRUNCATED) | [2.5744190216064453,1.5112360715866089,2.866237163543701,2.2959728240966797,1.470143437385559,1.6009(...TRUNCATED) |
ct3_T4_s4_contingent | 3 | 4 | 12 | 4 | contingent | true | null | 512 | 512 | 6 | 13 | [0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2,0,2(...TRUNCATED) | [5.096376419067383,0.636511504650116,4.366934299468994,10.524028778076172,4.956380367279053,9.597261(...TRUNCATED) | [-0.45175275206565857,2.7007546424865723,2.353661060333252,2.6778411865234375,2.2614777088165283,2.8(...TRUNCATED) |
ct3_T4_s4_shuffle_control | 3 | 4 | 12 | 4 | shuffle_control | false | 0.40929 | 520 | 504 | 0 | 0 | [2,0,0,2,0,2,0,2,0,0,2,2,2,2,2,0,2,0,2,0,2,0,2,0,2,2,0,0,0,2,2,0,2,0,2,0,0,2,2,0,0,2,2,0,2,0,2,0,2,0(...TRUNCATED) | [5.732888221740723,10.083234786987305,2.906954050064087,1.9007744789123535,0.2525610625743866,14.301(...TRUNCATED) | [2.7007546424865723,1.5702247619628906,0.6422613859176636,2.6778411865234375,2.660335063934326,2.858(...TRUNCATED) |
gallistel-streams
A fixed benchmark of synthetic Pavlovian conditioning streams for reading sequence models out
against an exact Bayes observer, generated with
gallistel-icl (scripts/make_benchmark.py).
This is an instrument, not animal data. Every stream is simulated. animal_law.json is
the only empirical content: slopes and timescale-invariance statements extracted from the
published animal literature (Harris & Gallistel 2026 rats; Gibbon & Balsam 1981 pigeons; Ward
et al. 2012 mice), with their citations, used as targets.
Contents
gallistel_streams.parquet(8.3 MB, zstd) andgallistel_streams.npz(same arrays): 960 streams = C/T in {3, 4.5, 6, 9, 13.5, 27, 54, 108} x T in {4, 12, 36} x 20 seeds x {contingent, shuffle_control}, 1024 tokens each.metadata.json: grid, seeding, column descriptions.animal_law.json: the animal slopes table with source citations.
Columns: stream_id, ct, T, C, seed, arm, contingent, shift_frac (control only),
n_cs_on, n_us, bayes_acq_full, bayes_acq_truncated, types (list of int8: 0 = CS_ON,
2 = US; 1 = CS_OFF is never emitted), dt (list of float32, seconds since the previous event),
target_logwait (list of float32, log seconds to the next US at every token: the model target).
Streams: inter-trial interval ~ Exp(C - T), CS-US interval ~ Exp(T), so the observer's rate
evidence per reinforced trial is log(C/T). Seeding is
np.random.default_rng([20261113, ct_index, t_index, seed]) then paired_streams(rng, CT, T),
so any stream can be regenerated from the package alone.
bayes_acq_full is the exact observer's acquisition trial (posterior >= 0.8 on 5 consecutive
CS_ON trials, prior 0.5) with the full LLR log(C/T) - w(1/T - 1/C); bayes_acq_truncated
uses the research code's truncated log(C/T) - w/T. 0 means never within the stream.
The shuffle-control leak (read before using the control as a null)
The control keeps the same CS_ON and US times with the CS process circularly shifted by
shift_frac of the horizon (uniform 0.2-0.8). Marginal rates are identical and the CS -> US
dependence is destroyed, which is what a rate-only null needs. But the shift also changes the
US -> CS_ON gap marginal (Exp(C - T) after every US in the contingent arm; arbitrary in the
control) and lets a CS_ON be followed by another CS_ON before any US. Interval shape and
alternation structure therefore separate the arms, not only rate. A learner can detect
contingency from that shape, and the rate-only observer saturates at one trial for C/T >= 15
on these streams. This is why no acquisition slope vs C/T could be measured in the research
run, and it is not fixed here. A control with a matched US -> CS_ON gap marginal, or a
full-likelihood observer that credits alternation evidence, is the needed next instrument.
Load
import pandas as pd
df = pd.read_parquet("hf://datasets/rohitravi2/gallistel-streams/gallistel_streams.parquet")
import numpy as np
z = np.load("gallistel_streams.npz") # z["types"] (960, 1024) int8, z["dt"], z["target_logwait"], z["ct"], ...
Model input encoding used by the released checkpoints: types as int64, log1p(dt) as float32.
Links
Code github.com/rohit-ravi2/gallistel-icl; weights rohitravi2/gallistel-icl; browser demo Space; write-up rohitravi.com/projects/ann-bnn-inference-law.
Citation
Ravi, R. (2026). gallistel-icl: in-context contingency inference read out against an exact Bayes observer on Gallistel-style conditioning streams. https://github.com/rohit-ravi2/gallistel-icl
- Downloads last month
- 41