File size: 7,926 Bytes
2bdd144
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
---
pretty_name: "Inductive Forecasting Study — Anonymous Data Release"
language:
- en
license: other
task_categories:
- text-generation
- reinforcement-learning
tags:
- forecasting
- reasoning
- prediction-markets
- synthetic-data
- behavioral-evaluation
configs:
- config_name: exp1_updates
  data_files:
  - split: test
    path: data/exp1/updates.parquet
- config_name: exp1_review_materials
  data_files:
  - split: test
    path: data/exp1/review_materials.parquet
- config_name: exp2_coin_city_tasks
  data_files:
  - split: test
    path: data/exp2/tasks.parquet
- config_name: exp2_coin_city_responses
  data_files:
  - split: test
    path: data/exp2/responses.parquet
- config_name: exp3_coin_city_transfer
  data_files:
  - split: train_causal
    path: data/exp3_coin_city/train_causal.parquet
  - split: train_population_prior
    path: data/exp3_coin_city/train_population_prior.parquet
  - split: train_structureless
    path: data/exp3_coin_city/train_structureless.parquet
  - split: test
    path: data/exp3_coin_city/test.parquet
- config_name: exp3_coin_city_scores
  data_files:
  - split: test
    path: data/exp3_coin_city/scores.parquet
- config_name: exp4_historical_markets
  data_files:
  - split: train
    path: data/exp4/train.parquet
  - split: validation
    path: data/exp4/validation.parquet
  - split: test
    path: data/exp4/test.parquet
- config_name: exp4_locked_test_outputs
  data_files:
  - split: test
    path: data/exp4/locked_test_outputs.parquet
- config_name: exp4_evidence_updates
  data_files:
  - split: test
    path: data/exp4/evidence_updates.parquet
- config_name: appendix_domain_transfer
  data_files:
  - split: train_d1
    path: data/appendix_domain/train_d1.parquet
  - split: train_d2
    path: data/appendix_domain/train_d2.parquet
  - split: train_d3
    path: data/appendix_domain/train_d3.parquet
  - split: train_structureless
    path: data/appendix_domain/train_structureless.parquet
  - split: test
    path: data/appendix_domain/test.parquet
- config_name: appendix_dag_transfer
  data_files:
  - split: train_family
    path: data/appendix_dag/train_family.parquet
  - split: train_population_prior
    path: data/appendix_dag/train_population_prior.parquet
  - split: train_structureless
    path: data/appendix_dag/train_structureless.parquet
  - split: test_family
    path: data/appendix_dag/test_family.parquet
  - split: test_population_prior
    path: data/appendix_dag/test_population_prior.parquet
- config_name: appendix_dag_zero_shot
  data_files:
  - split: test
    path: data/appendix_dag/zero_shot_scores.parquet
- config_name: appendix_mechanism_transfer
  data_files:
  - split: train
    path: data/appendix_mechanism/train_*.parquet
  - split: test
    path: data/appendix_mechanism/test_*.parquet
- config_name: appendix_mechanism_scores
  data_files:
  - split: test
    path: data/appendix_mechanism/scores.parquet
---

# Inductive Forecasting Study — Anonymous Data Release

This repository is the anonymous data companion to a paper studying behavioral
signatures of inductive reasoning in language-model forecasts. It packages the
frozen inputs, model responses, row-level scores, and aggregate result artifacts
used by the paper's four main experiments, together with synthetic appendix
transfer studies.

The release is organized as Hugging Face dataset configurations so each study can
be loaded independently:

```python
from datasets import load_dataset

updates = load_dataset("od2961/inductive-forecasting-data", "exp1_updates")
coin_city = load_dataset(
    "od2961/inductive-forecasting-data",
    "exp2_coin_city_responses",
)
historical = load_dataset(
    "od2961/inductive-forecasting-data",
    "exp4_historical_markets",
)
```

## Contents

### Experiment 1: selective updating

`exp1_updates` contains one row per frozen model × market × evidence packet ×
run record from the June 17 numerical authority. It includes initial and updated
probabilities, evidence direction, EHC/HFC/ICS indicators, and parsing status.
The source report and threshold analysis are under `artifacts/exp1/`.

`exp1_review_materials` contains the 18 blinded fictional reports shown in the
human materials review. Individual participant ratings and timestamps are not
released because participation consent did not explicitly establish consent for
public row-level redistribution. Aggregate review results remain available under
`artifacts/exp1/`.

### Experiment 2: Coin City

`exp2_coin_city_tasks` contains the frozen target-only, no-context,
correct-context, misleading-context, and arbitrary-symbol prompts joined to their
answer key. `exp2_coin_city_responses` contains all production model responses,
including the matched arbitrary-symbol control set, joined to the same task and
gold fields. Failed pilots and corrupt pre-repair files are excluded.

### Experiment 3: trained Coin City transfer

`exp3_coin_city_transfer` contains the 4,800-row causal, population-prior, and
structureless training sets plus the common 1,440-row held-out evaluation set.
`exp3_coin_city_scores` contains the registered greedy and five-draw Qwen3-4B
endpoint score rows for the base, causal, population-prior, and structureless
conditions. Checkpoints are not part of this dataset repository.

### Experiment 4: historical-market transfer

`exp4_historical_markets` contains the checksum-locked 1,736/512/1,024
train/development/test tasks derived from a public Polymarket archive.
`exp4_locked_test_outputs` contains base and three-seed final forecasts.
`exp4_evidence_updates` contains the secondary paired evidence-update outputs.
The 21 GB upstream scrape is not redistributed.

### Appendix studies

The remaining configurations contain the registered Coin-* domain ladder, DAG
family, and mechanism-composition train/evaluation datasets. Final score ledgers
are included when the paper's frozen analysis manifest names an exact ledger.

## Provenance and construction

- Real-market questions and historical prices were collected from Polymarket's
  public interfaces. The derived task rows preserve the study's point-in-time
  filtering and family-disjoint split.
- Coin City, Coin-*, DAG, and mechanism-family tasks are synthetic and generated
  by the study code.
- Model response rows are machine-generated outputs from the deployments named in
  each record or accompanying artifact.
- `release_manifest.json` records row counts and explicit exclusions.
- `SHA256SUMS` authenticates every published file.

Absolute cluster paths and author-identifying endpoint names are not included.
Authentication secrets were never intentionally written to the source records;
the release builder also performs a fail-closed secret and identity scan.

## Limitations and responsible use

The synthetic studies are behavioral benchmarks, not demonstrations of a
particular internal representation. The historical-market tasks inherit selection
biases and limitations of prediction-market archives. Static question/rules text
cannot always be proven point-in-time, as described in the paper. Model outputs
may contain errors or unsupported claims and should not be treated as factual or
as financial advice.

Some Experiment 1 records contain fictional, explicitly labeled evidence packets
about real people and events. Preserve the fictional notice when displaying or
redistributing these materials.

## Licensing

See `LICENSE.md`. The repository uses `license: other` because it combines
original synthetic benchmarks with third-party market-derived fields and model
outputs that are subject to their respective source/provider terms. No blanket
relicensing of third-party material is asserted.

## Citation

During double-blind review, cite this repository as:

> Anonymous Authors (2026). *Inductive Forecasting Study — Anonymous Data
> Release*. Hugging Face Datasets.

The card and citation should be updated after deanonymization.