Mental Spaces Corpus v0.1

.gitattributes

@@ -0,0 +1,60 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.avro filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.lz4 filter=lfs diff=lfs merge=lfs -text
+*.mds filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+# Audio files - uncompressed
+*.pcm filter=lfs diff=lfs merge=lfs -text
+*.sam filter=lfs diff=lfs merge=lfs -text
+*.raw filter=lfs diff=lfs merge=lfs -text
+# Audio files - compressed
+*.aac filter=lfs diff=lfs merge=lfs -text
+*.flac filter=lfs diff=lfs merge=lfs -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text
+*.ogg filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+# Image files - uncompressed
+*.bmp filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.tiff filter=lfs diff=lfs merge=lfs -text
+# Image files - compressed
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text
+*.webp filter=lfs diff=lfs merge=lfs -text
+# Video files - compressed
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.webm filter=lfs diff=lfs merge=lfs -text

CITATION.cff

@@ -0,0 +1,38 @@
+cff-version: 1.2.0
+message: "If you use this dataset, please cite it using this metadata and the Hugging Face DOI."
+title: "Mental Spaces Corpus"
+version: "0.1.0"
+type: dataset
+authors:
+  - family-names: Steele
+    given-names: Oliver
+license: MIT
+repository-code: "https://github.com/osteele/mental-spaces"
+abstract: >-
+  A controlled suite of natural-language stimuli for testing whether language
+  models keep base-space and alternative-space discourse targets separate under
+  mental-space constructions such as counterfactuals, belief contexts,
+  depictive spaces, and nested spaces.
+references:
+  - type: article
+    title: "One mechanism for many mental spaces: a shared router over a value slot in language models"
+    authors:
+      - family-names: Steele
+        given-names: Oliver
+      - family-names: Wen
+        given-names: Jiangtao
+      - family-names: Han
+        given-names: Yuxing
+    year: 2026
+    url: "https://arxiv.org/abs/2607.10248"
+  - type: article
+    title: "Belief-reality separation lives in routing over a shared value slot in language models"
+    authors:
+      - family-names: Steele
+        given-names: Oliver
+      - family-names: Wen
+        given-names: Jiangtao
+      - family-names: Han
+        given-names: Yuxing
+    year: 2026
+    url: "https://arxiv.org/abs/2607.11945"

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Oliver Steele
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,171 @@
+---
+pretty_name: Mental Spaces Corpus
+language:
+- en
+license: mit
+task_categories:
+- text-generation
+- text-classification
+tags:
+- mental-spaces
+- counterfactual-reasoning
+- belief-tracking
+- mechanistic-interpretability
+- minimal-pairs
+- text
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/train.jsonl
+  - split: validation
+    path: data/validation.jsonl
+  - split: test
+    path: data/test.jsonl
+---
+
+# Mental Spaces Corpus
+
+Version: 0.1.0
+
+The Mental Spaces Corpus is a controlled suite of natural-language stimuli for testing
+whether language models keep base-space and alternative-space discourse targets
+separate. It is designed for probing, causal interventions, and behavioral readouts in
+mental-space constructions such as counterfactuals, belief contexts, and depictive
+spaces, including nested belief and nested depictive spaces.
+
+This release is a stimulus suite for controlled experiments, not a broad natural-language
+benchmark. It provides confound control: items expose explicit base-space and
+alternative-space target values, matched readout prompts, and annotations for the space
+builder and control type.
+
+Each construction is realized across several surface *frames* (distinct templates, tracked
+by `template_id`) so that results do not rest on a single fixed sentence. Counterfactual
+frames vary the space builder itself (painting / dream / movie / story / hypothetical);
+every other construction holds the space role constant and varies only the wording.
+
+## Dataset Structure
+
+Files:
+
+- `data/train.jsonl`
+- `data/validation.jsonl`
+- `data/test.jsonl`
+- `schema.json`
+- `summary.json`
+- `VALIDATION.md`
+- `CITATION.cff`
+- `LICENSE`
+
+Each JSONL row is one readout prompt. A single underlying scenario can yield multiple
+rows, for example a reality readout and a belief readout.
+
+Core fields:
+
+- `context`: the full setup text.
+- `prompt`: the completion prompt ending immediately before the expected answer.
+- `answer`: the expected color/value continuation.
+- `scenario_id`: stable ID shared by all readout rows from one scenario.
+- `construction`: the construction family, such as `belief_reality` or
+  `counterfactual_or_alternative_space`.
+- `answer_space`: the discourse space queried by the prompt.
+- `base_value` and `alternative_value`: the competing target values when present.
+- `value_space_map`: all target values keyed by space, or by entity-and-space for
+  multi-entity controls.
+- `template_id`: stable template/frame identifier for frame-sensitive analyses.
+- `answer_kind`: the answer vocabulary type, currently `color`.
+- `confound_controls`: design controls active for the item.
+- `source_experiment`: project experiment family the construction derives from.
+
+See `schema.json` for the complete field list.
+
+## Intended Use
+
+The corpus is intended for:
+
+- behavioral completion tests;
+- linear probes over model activations;
+- activation patching and low-rank steering experiments;
+- controlled comparisons of base-space, alternative-space, and nested-space readouts.
+
+Prompts are completion-style and currently use short color/value answers. Users should
+verify tokenizer behavior for a target model before interpreting token-level results.
+
+## Design Controls and Known Confounds
+
+This corpus is built around known failure modes in counterfactual, belief-tracking, and
+theory-of-mind evaluations. Each row includes a `confound_controls` list that records
+which controls are active for that item.
+
+The main controls are:
+
+- `distinct_values`: competing spaces use different answer values, so a model must
+  choose a space rather than copy the only value in the context.
+- `randomized_clause_order`: base-space and alternative-space clauses appear in
+  randomized order, reducing first/last-mention and recency shortcuts.
+- `generic_entities`: counterfactual and depictive items use arbitrary entities and
+  colors, following the generic-subject logic of Li, Yu, and Ettinger (2023), so world
+  knowledge is not enough to answer.
+- `copy_proof_readout`: belief and nested-belief rows include both relevant values in
+  context, addressing the copying concern raised by theory-of-mind benchmark critiques
+  such as Ullman (2023), Sclar et al. (2023), and Shapira et al. (2024).
+- `two_entities` and `entity_bound_readout`: two-entity rows distinguish a global
+  image-versus-reality feature from an entity-bound space assignment.
+- `three_distinct_values` and `nested_space_builder`: nested depictive rows separate
+  reality, inner-depiction, and outer-depiction values (for example a photograph and a
+  painting of that photograph) so nested-space behavior cannot be reduced to a binary
+  frame choice.
+
+These controls do not make the corpus a solved benchmark. They are meant to make the
+stimuli inspectable and to expose likely shortcuts. Users should still report
+tokenization checks, clean behavioral accuracy, and whether results survive irrelevant
+context or frame variation. `VALIDATION.md` reports a behavioral check of the released
+stimuli on Qwen2.5-7B-Instruct. The context-fragility concern follows Schouten et al.
+(2024), and the broader requirement that a representation be accurate, coherent,
+uniform, and causally used follows Herrmann and Levinstein (2024).
+
+## Related Papers
+
+These constructions are studied in:
+
+- Steele, Oliver. 2026. One mechanism for many mental spaces: a shared router over a value
+  slot in language models. arXiv:2607.10248.
+- Steele, Oliver. 2026. Belief-reality separation lives in routing over a shared value slot
+  in language models. arXiv:2607.11945.
+
+The corpus is a companion stimulus suite for the mental-space phenomena these papers study,
+not the exact experimental data behind their reported numbers.
+
+## References
+
+- Herrmann, Daniel A., and Benjamin A. Levinstein. 2024. Criteria for belief-like
+  representations in language models.
+- Li, Jiaxuan, Lang Yu, and Allyson Ettinger. 2023. Counterfactual reasoning in
+  language models.
+- Sclar, Melanie, Sachin Kumar, Peter West, Alane Suhr, Yejin Choi, and Yulia
+  Tsvetkov. 2023. Benchmark concerns for theory-of-mind evaluation.
+- Schouten et al. 2024. Belief probes are fragile under irrelevant context.
+- Shapira, Natalie, et al. 2024. Clever Hans-style shortcuts in theory-of-mind
+  evaluation.
+- Ullman, Tomer. 2023. Large language models fail on trivial alterations to
+  theory-of-mind tasks.
+
+## Citation
+
+Please cite this dataset via `CITATION.cff` (Oliver Steele) and
+the Hugging Face dataset DOI. The dataset lives at
+https://huggingface.co/datasets/osteele/mental-spaces.
+
+## Provenance
+
+The templates derive from the `mental-spaces` project experiment suite; the release adds
+surface-frame variants (further space builders and wordings) to broaden coverage beyond
+the single frame each experiment used. The build script lives in the source repository,
+[github.com/osteele/mental-spaces](https://github.com/osteele/mental-spaces); the generated
+files are deterministic:
+
+```bash
+uv run python scripts/build_corpus.py --profile expanded
+```
+
+The checked-in `summary.json` states which profile generated the current files.

VALIDATION.md

@@ -0,0 +1,112 @@
+# Behavioral Validation
+
+This records whether the released stimuli actually elicit the intended space-tracking
+behavior in capable open models. It is a check on the stimuli, not a reproduction of any
+result in the papers that study these constructions (arXiv:2607.10248, arXiv:2607.11945):
+those papers' numbers come from the project experiment scripts, and this corpus is a
+companion stimulus suite derived from the same constructions.
+
+## Method
+
+Each row's `prompt` is a completion that ends on a copula, so the answer is the next
+token. For every prompt we read the next-token logits and take:
+
+- **forced accuracy** (primary): argmax restricted to the item's in-play values
+  (`value_space_map`). This is the minimal-pair question: does the model pick the correct
+  space's value over the competing one?
+- **open accuracy** (secondary): argmax over the full color vocabulary. The gap between
+  open and forced surfaces the model's color prior and any off-vocabulary drift; it is a
+  diagnostic, not the headline.
+
+Answers are scored on the first token of ` {color}` using the target model's own
+tokenizer; the single-token audit below reports whether each answer is one token.
+
+Scoring is completion-style (no chat template), matching the corpus format. Instruction
+models may behave differently under their chat template; verify for your setup.
+
+## Caveats
+
+- **Two architectures, still provisional.** These numbers are Qwen2.5-7B-Instruct and
+  OLMo-2-1124-7B-Instruct. Agreement across two models is reassuring but not proof of
+  generality. The cross-model pattern matters more than any single cell: a frame weak on
+  one model may be fine on another, and only frames weak on both point to the stimulus.
+- Accuracy below 1.0 is expected: two-space and nested constructions are genuinely hard,
+  and the two-entity control is designed to leak (see below). The bar is that a capable
+  model tracks the correct space well above chance, not perfectly.
+
+## Per-construction results
+
+Completion scoring, all splits, forced accuracy (argmax restricted to the item's competing
+values). Overall 0.87 (Qwen2.5-7B-Instruct) and 0.90 (OLMo-2-1124-7B-Instruct).
+
+| construction | n | Qwen-7B | OLMo-2-7B |
+|---|---:|---:|---:|
+| single_space_control | 120 | 1.000 | 1.000 |
+| belief_reality | 450 | 0.933 | 0.933 |
+| counterfactual_or_alternative_space | 500 | 0.846 | 0.930 |
+| nested_depictive | 360 | 0.825 | 0.892 |
+| two_entity_selectivity | 240 | 0.858 | 0.854 |
+| nested_belief | 240 | 0.821 | 0.817 |
+
+Both models track the correct space well above the two-value chance line (0.50) on every
+construction. Forced accuracy is the metric that answers the space-tracking question; the
+open six-way argmax deflates the four-color constructions through the model's prior toward
+the two unused colors, so it stays a diagnostic rather than the headline.
+
+The two-entity control leaks as designed (swap 0.125 Qwen, 0.142 OLMo-2): it separates
+entity-bound space assignment from a global image-versus-reality feature, and a model that
+copies the distractor's value scores as swap. Nested belief is the hardest construction on
+both models, driven by one frame (see below).
+
+## Per-template results
+
+Forced accuracy per frame. This is the diagnostic for a weak surface realization.
+
+| construction | template | n | Qwen-7B | OLMo-2-7B |
+|---|---|---:|---:|---:|
+| belief_reality | belief-thinks | 150 | 0.967 | 0.860 |
+| belief_reality | belief-convinced | 150 | 0.960 | 0.960 |
+| belief_reality | belief-believes | 150 | 0.873 | 0.980 |
+| counterfactual | cf-hypothetical | 100 | 0.940 | 0.940 |
+| counterfactual | cf-painting | 100 | 0.810 | 0.910 |
+| counterfactual | cf-story | 100 | 0.910 | 0.900 |
+| counterfactual | cf-dream | 100 | 0.800 | 0.930 |
+| counterfactual | cf-movie | 100 | 0.770 | 0.970 |
+| nested_belief | nb-imagines-confirms | 80 | 0.975 | 0.900 |
+| nested_belief | nb-assumes-sees | 80 | 0.800 | 0.838 |
+| nested_belief | nb-thinks-knows | 80 | 0.688 | 0.713 |
+| nested_depictive | nd-reflection-portrait | 120 | 0.867 | 0.933 |
+| nested_depictive | nd-photo-painting | 120 | 0.825 | 0.900 |
+| nested_depictive | nd-snapshot-sketch | 120 | 0.783 | 0.842 |
+| single_space_control | single-space-* | 40 each | 1.000 | 1.000 |
+| two_entity_selectivity | te-painting | 80 | 0.850 | 0.912 |
+| two_entity_selectivity | te-photo | 80 | 0.863 | 0.838 |
+| two_entity_selectivity | te-drawing | 80 | 0.863 | 0.812 |
+
+`nb-thinks-knows` is the hardest nested-belief frame (0.69 Qwen, 0.71 OLMo-2) but no longer
+near chance. An earlier build read 0.60 / 0.55: the readout then repeated a verb
+(`{observer} thinks {believer} thinks ...`), which let the model collapse the two belief
+levels. Changing the outer verb to `believes` lifted the frame on both models (+0.09 Qwen,
++0.16 OLMo-2), so the cause was surface phrasing, not a defect in the scenario. It stays the
+softest nested-belief frame, consistent with genuine 2-deep difficulty on top of the
+phrasing effect.
+
+Frame variance elsewhere is model-specific rather than stimulus-driven: `cf-movie` is Qwen's
+softest counterfactual (0.77) but OLMo-2's strongest (0.97).
+
+## Single-token audit
+
+All answer colors are a single token under both models' tokenizers (Qwen2.5-7B-Instruct and
+OLMo-2-1124-7B-Instruct), with the leading space: blue, green, red, brown, yellow, purple.
+Re-check for any other model before interpreting token-level results.
+
+## Reproduce
+
+The build and validation scripts live in the source repository,
+[github.com/osteele/mental-spaces](https://github.com/osteele/mental-spaces). With the
+model and a GPU available, the validation script scores every prompt (`uv run` installs
+`torch` and `transformers` from the lockfile):
+
+```bash
+uv run python scripts/exp_corpus_behavioral_audit.py --model Qwen/Qwen2.5-7B-Instruct
+```

schema.json

@@ -0,0 +1,120 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "additionalProperties": false,
+  "properties": {
+    "alternative_space": {
+      "type": ["string", "null"]
+    },
+    "alternative_value": {
+      "type": ["string", "null"]
+    },
+    "answer": {
+      "type": "string"
+    },
+    "answer_kind": {
+      "type": "string"
+    },
+    "answer_space": {
+      "type": "string"
+    },
+    "base_space": {
+      "type": "string"
+    },
+    "base_value": {
+      "type": ["string", "null"]
+    },
+    "confound_controls": {
+      "items": {
+        "type": "string"
+      },
+      "type": "array"
+    },
+    "construction": {
+      "type": "string"
+    },
+    "context": {
+      "type": "string"
+    },
+    "control_type": {
+      "type": "string"
+    },
+    "entity": {
+      "type": "string"
+    },
+    "family": {
+      "type": "string"
+    },
+    "id": {
+      "type": "string"
+    },
+    "labels": {
+      "additionalProperties": {
+        "type": ["string", "boolean"]
+      },
+      "type": "object"
+    },
+    "lexical_trigger_present": {
+      "type": "boolean"
+    },
+    "notes": {
+      "type": "string"
+    },
+    "prompt": {
+      "type": "string"
+    },
+    "query_type": {
+      "type": "string"
+    },
+    "scenario_id": {
+      "type": "string"
+    },
+    "source_experiment": {
+      "type": "string"
+    },
+    "space_builder": {
+      "type": ["string", "null"]
+    },
+    "template_id": {
+      "type": "string"
+    },
+    "value_space_map": {
+      "additionalProperties": {
+        "type": "string"
+      },
+      "type": "object"
+    },
+    "split": {
+      "enum": ["train", "validation", "test"],
+      "type": "string"
+    }
+  },
+  "required": [
+    "id",
+    "scenario_id",
+    "split",
+    "construction",
+    "family",
+    "source_experiment",
+    "context",
+    "prompt",
+    "answer",
+    "answer_space",
+    "entity",
+    "base_space",
+    "alternative_space",
+    "base_value",
+    "alternative_value",
+    "value_space_map",
+    "query_type",
+    "template_id",
+    "answer_kind",
+    "space_builder",
+    "lexical_trigger_present",
+    "control_type",
+    "labels",
+    "confound_controls",
+    "notes"
+  ],
+  "title": "Mental Spaces Corpus record",
+  "type": "object"
+}

summary.json

@@ -0,0 +1,53 @@
+{
+  "constructions": [
+    "belief_reality",
+    "counterfactual_or_alternative_space",
+    "nested_belief",
+    "nested_depictive",
+    "single_space_control",
+    "two_entity_selectivity"
+  ],
+  "families": [
+    "belief",
+    "controls",
+    "counterfactual",
+    "depictive"
+  ],
+  "profile": "expanded",
+  "records": 1910,
+  "source_experiments": [
+    "EXP-001",
+    "EXP-001/002/003",
+    "EXP-030",
+    "EXP-034",
+    "EXP-052",
+    "EXP-067"
+  ],
+  "splits": {
+    "test": 284,
+    "train": 1337,
+    "validation": 289
+  },
+  "templates": [
+    "belief-believes",
+    "belief-convinced",
+    "belief-thinks",
+    "cf-dream",
+    "cf-hypothetical",
+    "cf-movie",
+    "cf-painting",
+    "cf-story",
+    "nb-assumes-sees",
+    "nb-imagines-confirms",
+    "nb-thinks-knows",
+    "nd-photo-painting",
+    "nd-reflection-portrait",
+    "nd-snapshot-sketch",
+    "single-space-color-of",
+    "single-space-plain",
+    "single-space-there-is",
+    "te-drawing",
+    "te-painting",
+    "te-photo"
+  ]
+}