Datasets:

awni00
/

multi-strategy-algorithmic-tasks

+---
+pretty_name: Multi-Strategy Algorithmic Tasks
+license: cc-by-4.0
+language:
+  - en
+task_categories:
+  - text-generation
+tags:
+  - synthetic
+  - reasoning
+  - algorithmic-reasoning
+  - strategy
+  - datasets
+size_categories:
+  - 1M<n<10M
+---
+# Multi-Strategy Algorithmic Tasks
+Multi-Strategy Algorithmic Tasks is a synthetic benchmark of parseable
+algorithmic problems with multiple valid solution strategies. Each example
+contains a problem, a complete strategy-specific solution trace, and the
+strategy used to generate that trace.
+The benchmark accompanies
+[Uncovering Latent Reasoning Strategies in Language Models](https://arxiv.org/abs/2607.17674).
+It provides controlled reference strategy labels for studying strategy
+recovery, representation, routing, and controllable generation.
+## Load the dataset
+```python
+from datasets import load_dataset
+dataset = load_dataset("awni00/multi-strategy-algorithmic-tasks")
+example = dataset["train"][0]
+```
+The release contains:
+| Split | Rows |
+| --- | ---: |
+| `train` | 1,000,000 |
+| `validation` | 10,000 |
+| `test` | 10,000 |
+To select one task family:
+```python
+sorting = dataset.filter(
+    lambda example: example["task_name"] == "sorting_algorithms"
+)
+```
+## Fields
+All four fields are strings.
+| Field | Description |
+| --- | --- |
+| `task_name` | Algorithmic task family |
+| `strategy_id` | Namespaced strategy sampled to generate the trace |
+| `input_text` | Rendered problem instance |
+| `reasoning_trace` | Complete strategy-specific solution trace, including the final answer |
+The dataset intentionally contains text rather than token IDs. Users can apply
+the tokenizer and sequence framing appropriate for their own model.
+Example:
+```python
+{
+    "task_name": "multidigit_addition",
+    "strategy_id": "multidigit_addition:left-to-right-partials",
+    "input_text": "560+342",
+    "reasoning_trace": (
+        "p100:500+300=800 ; p10:60+40=100 ; "
+        "p1:0+2=2 ; sum=800+100+2=902"
+    ),
+}
+```
+## Tasks and strategies
+The generator first samples one of the six task families uniformly. It then
+samples a strategy uniformly within that task family.
+| Task | Problem | Strategies |
+| --- | --- | --- |
+| `list_summation` | Sum four integers from 0 to 9 | `left-to-right`, `right-to-left`, `pairwise` |
+| `sorting_algorithms` | Sort five integers from 0 to 9 | `bubble-sort`, `selection-sort`, `insertion-sort`, `merge-sort`, `heap-sort` |
+| `grid_pathfinding` | Monotone shortest paths on a 6×6 grid | `right-first`, `down-first`, `alternating` |
+| `linear_equation_solving` | Solve integer equations of the form `ax+b=c` | `subtract-then-divide`, `divide-then-subtract`, `inverse-ops` |
+| `base_conversion` | Convert integers from 1 to 255 to base 2, 4, 8, or 16 | `repeated-division`, `via-binary`, `decomposition` |
+| `multidigit_addition` | Add two three-digit nonnegative integers | `right-to-left-carry`, `left-to-right-partials`, `rounding-decomposition` |
+The public `strategy_id` includes the task namespace, for example
+`grid_pathfinding:alternating`. There are 20 namespaced strategies in total.
+## Generation and reproducibility
+The published rows are a text rendering of the frozen aggregate dataset used
+for the accompanying paper. The release preserves the original examples,
+split membership, and row order. It does not resample or filter the data.
+The task family, input, and strategy are sampled as:
+\[
+T \sim \operatorname{Unif}(\text{task families}), \qquad
+X \sim \mathcal{D}_T, \qquad
+S \sim \operatorname{Unif}(\mathcal{S}_T), \qquad
+Y = \operatorname{Trace}_T(X,S).
+\]
+The repository includes `generation_config.yaml` with the resolved task
+parameters and random seeds. `release_manifest.json` records source and
+release checksums. `validation_report.json` records task and strategy counts,
+trace ambiguity, duplicates, and cross-split overlap.
+## Important characteristics
+### Sampled strategy labels
+`strategy_id` is the strategy selected by the generator. On some inputs,
+multiple strategies produce the same observable trace. Such a trace is valid
+but its sampled strategy may not be uniquely recoverable from the trace alone.
+This is most common when a problem requires only a few steps or different
+algorithms happen to traverse identical intermediate states.
+### Sampling with replacement
+Examples are sampled with replacement from finite task distributions.
+Repeated examples within a split and overlap between splits are therefore
+expected. The splits represent independent samples from the same controlled
+distribution; they are not disjoint-input generalization splits.
+## Intended use
+The dataset is intended for controlled research on:
+- learning and evaluating multiple reasoning strategies;
+- strategy recovery from generated traces or model representations;
+- latent-variable routing and controllable generation;
+- algorithmic sequence modeling with parseable outputs.
+The reference strategy labels are intended for evaluation and analysis. In the
+paper's main setup, they are not provided as supervision to the language model.
+## Limitations
+- The tasks are synthetic, symbolic, and deliberately small.
+- Difficulty ranges are fixed by the published generation configuration.
+- The traces implement a finite reference set of procedures, not every valid
+  solution method.
+- Strategy labels describe the generator procedure and should not be treated
+  as natural human reasoning annotations.
+- Results on this benchmark do not by themselves establish strategy recovery
+  on open-ended natural-language reasoning.
+The dataset contains no human-authored examples or personal information.
+## License
+The dataset is released under the
+[Creative Commons Attribution 4.0 International License](https://creativecommons.org/licenses/by/4.0/).
+## Citation
+```bibtex
+@misc{altabaa2026uncovering,
+  title         = {Uncovering Latent Reasoning Strategies in Language Models},
+  author        = {Awni Altabaa and John Lafferty},
+  year          = {2026},
+  eprint        = {2607.17674},
+  archivePrefix = {arXiv},
+  primaryClass  = {cs.LG},
+  url           = {https://arxiv.org/abs/2607.17674}
+}
+```
+Code associated with the paper is available at
+[Awni00/latent-strategies-in-lms](https://github.com/Awni00/latent-strategies-in-lms).

data/test.parquet ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:9c322f640f5e8f6ef448f87281340252df0c75b3fd45175ab413530cad484d41
+size 205404

data/train.parquet ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:03563441091e962868c4f558e2e9b22aae3970cba7e53d56aa1707d7899e5c35
+size 19837251

data/validation.parquet ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:3a4d2d70b9640ecddc088edc12320417fd2d2314a0bb656e6f0feca405c57636
+size 205088

generation_config.yaml ADDED Viewed

	@@ -0,0 +1,50 @@

+seed: 0
+split_sizes:
+  train: 1_000_000
+  val: 10_000
+  test: 10_000
+split_seed_offsets:
+  train: 0
+  val: 1_000_000
+  test: 2_000_000
+task_seed_stride: 100_000
+include_semantic: false
+strict_validation: false
+data_dir: .data/synthetic_sequences
+task_args:
+  multi_task:
+    included_task_names:
+      - list_summation
+      - sorting_algorithms
+      - grid_pathfinding
+      - linear_equation_solving
+      - base_conversion
+      - multidigit_addition
+    subtask_args:
+      list_summation:
+        min_value: 0
+        max_value: 9
+        list_length: 4
+      sorting_algorithms:
+        min_value: 0
+        max_value: 9
+        list_length: 5
+      grid_pathfinding:
+        grid_size: 6
+      linear_equation_solving:
+        min_a: 2
+        max_a: 9
+        min_solution: -9
+        max_solution: 9
+        min_k: -9
+        max_k: 9
+      base_conversion:
+        min_value: 1
+        max_value: 255
+        allowed_bases:
+          - 2
+          - 4
+          - 8
+          - 16
+      multidigit_addition:
+        digits: 3

release_manifest.json ADDED Viewed

	@@ -0,0 +1,63 @@

+{
+  "dataset_id": "awni00/multi-strategy-algorithmic-tasks",
+  "public_schema": {
+    "input_text": "string",
+    "reasoning_trace": "string",
+    "strategy_id": "string",
+    "task_name": "string"
+  },
+  "release_files": {
+    "README": {
+      "path": "README.md",
+      "sha256": "2704f4a4840dd372acffd65f32144df875be0de9123222506d6d74dd17b0eb82"
+    },
+    "generation_config": {
+      "path": "generation_config.yaml",
+      "sha256": "945fecea0fc4b246aa08b106e59d6072c04ddb0fbd0e9b71740e47e5f2c943a0"
+    },
+    "test": {
+      "num_rows": 10000,
+      "path": "data/test.parquet",
+      "sha256": "9c322f640f5e8f6ef448f87281340252df0c75b3fd45175ab413530cad484d41"
+    },
+    "train": {
+      "num_rows": 1000000,
+      "path": "data/train.parquet",
+      "sha256": "03563441091e962868c4f558e2e9b22aae3970cba7e53d56aa1707d7899e5c35"
+    },
+    "validation": {
+      "num_rows": 10000,
+      "path": "data/validation.parquet",
+      "sha256": "3a4d2d70b9640ecddc088edc12320417fd2d2314a0bb656e6f0feca405c57636"
+    },
+    "validation_report": {
+      "path": "validation_report.json",
+      "sha256": "f0973a532e9ae165f064b7b35116a2e4eb5277784aac990727a46e560a3439e2"
+    }
+  },
+  "release_version": "1.0.0",
+  "source": {
+    "directory": ".data/synthetic_sequences/multi_task",
+    "files": {
+      "test": {
+        "num_rows": 10000,
+        "path": "test.parquet",
+        "sha256": "fd2c489e4b8fbea50526979dfe27ac5ddec9d6c8df0db9eb93117b39f6730785"
+      },
+      "train": {
+        "num_rows": 1000000,
+        "path": "train.parquet",
+        "sha256": "dd172fee55aa89d30a2fe53aa4f7e1b7fe3f0b4899879ed3e71b95b9cf3bc765"
+      },
+      "val": {
+        "num_rows": 10000,
+        "path": "val.parquet",
+        "sha256": "f4862c9a9a8f1ae2c02debfb7ca6ce79b134e479d945c7c22ba996206bfda290"
+      }
+    },
+    "generated_at": "2026-04-26T00:01:29",
+    "metadata_schema_version": "synthetic_sequences.task_metadata.v1",
+    "metadata_sha256": "24087732ffedb083b763be12c00786be1f7bb7c3df2869dc066b520aa2e22596",
+    "task_version": "1.0.0"
+  }
+}

validation_report.json ADDED Viewed

	@@ -0,0 +1,196 @@

+{
+  "cross_split_overlap": {
+    "train__test": {
+      "shared_unique_inputs": 5792,
+      "shared_unique_rows": 6209
+    },
+    "train__validation": {
+      "shared_unique_inputs": 5845,
+      "shared_unique_rows": 6254
+    },
+    "validation__test": {
+      "shared_unique_inputs": 1777,
+      "shared_unique_rows": 1356
+    }
+  },
+  "fingerprint_definition": {
+    "algorithm": "sha256-length-framed-utf8",
+    "input": [
+      "task_name",
+      "input_text"
+    ],
+    "row": [
+      "task_name",
+      "strategy_id",
+      "input_text",
+      "reasoning_trace"
+    ]
+  },
+  "schema": {
+    "input_text": "string",
+    "reasoning_trace": "string",
+    "strategy_id": "string",
+    "task_name": "string"
+  },
+  "splits": {
+    "test": {
+      "ambiguous_rate": 0.0662,
+      "ambiguous_rows": 662,
+      "ambiguous_task_counts": {
+        "grid_pathfinding": 497,
+        "sorting_algorithms": 165
+      },
+      "duplicate_rows": 1673,
+      "num_rows": 10000,
+      "public_split_name": "test",
+      "repeated_inputs": 2870,
+      "source_split_name": "test",
+      "strategy_counts": {
+        "base_conversion:decomposition": 570,
+        "base_conversion:repeated-division": 575,
+        "base_conversion:via-binary": 507,
+        "grid_pathfinding:alternating": 556,
+        "grid_pathfinding:down-first": 530,
+        "grid_pathfinding:right-first": 571,
+        "linear_equation_solving:divide-then-subtract": 552,
+        "linear_equation_solving:inverse-ops": 615,
+        "linear_equation_solving:subtract-then-divide": 555,
+        "list_summation:left-to-right": 540,
+        "list_summation:pairwise": 566,
+        "list_summation:right-to-left": 518,
+        "multidigit_addition:left-to-right-partials": 569,
+        "multidigit_addition:right-to-left-carry": 529,
+        "multidigit_addition:rounding-decomposition": 564,
+        "sorting_algorithms:bubble-sort": 320,
+        "sorting_algorithms:heap-sort": 352,
+        "sorting_algorithms:insertion-sort": 333,
+        "sorting_algorithms:merge-sort": 351,
+        "sorting_algorithms:selection-sort": 327
+      },
+      "task_counts": {
+        "base_conversion": 1652,
+        "grid_pathfinding": 1657,
+        "linear_equation_solving": 1722,
+        "list_summation": 1624,
+        "multidigit_addition": 1662,
+        "sorting_algorithms": 1683
+      },
+      "unique_inputs": 7130,
+      "unique_rows": 8327
+    },
+    "train": {
+      "ambiguous_rate": 0.066884,
+      "ambiguous_rows": 66884,
+      "ambiguous_task_counts": {
+        "grid_pathfinding": 51002,
+        "list_summation": 13,
+        "sorting_algorithms": 15869
+      },
+      "duplicate_rows": 696118,
+      "num_rows": 1000000,
+      "public_split_name": "train",
+      "repeated_inputs": 805381,
+      "source_split_name": "train",
+      "strategy_counts": {
+        "base_conversion:decomposition": 55857,
+        "base_conversion:repeated-division": 55377,
+        "base_conversion:via-binary": 55784,
+        "grid_pathfinding:alternating": 55490,
+        "grid_pathfinding:down-first": 55740,
+        "grid_pathfinding:right-first": 55647,
+        "linear_equation_solving:divide-then-subtract": 55737,
+        "linear_equation_solving:inverse-ops": 55601,
+        "linear_equation_solving:subtract-then-divide": 55350,
+        "list_summation:left-to-right": 55430,
+        "list_summation:pairwise": 55421,
+        "list_summation:right-to-left": 55171,
+        "multidigit_addition:left-to-right-partials": 55920,
+        "multidigit_addition:right-to-left-carry": 55284,
+        "multidigit_addition:rounding-decomposition": 55486,
+        "sorting_algorithms:bubble-sort": 33429,
+        "sorting_algorithms:heap-sort": 33399,
+        "sorting_algorithms:insertion-sort": 33391,
+        "sorting_algorithms:merge-sort": 33252,
+        "sorting_algorithms:selection-sort": 33234
+      },
+      "task_counts": {
+        "base_conversion": 167018,
+        "grid_pathfinding": 166877,
+        "linear_equation_solving": 166688,
+        "list_summation": 166022,
+        "multidigit_addition": 166690,
+        "sorting_algorithms": 166705
+      },
+      "unique_inputs": 194619,
+      "unique_rows": 303882
+    },
+    "validation": {
+      "ambiguous_rate": 0.0652,
+      "ambiguous_rows": 652,
+      "ambiguous_task_counts": {
+        "grid_pathfinding": 491,
+        "list_summation": 1,
+        "sorting_algorithms": 160
+      },
+      "duplicate_rows": 1682,
+      "num_rows": 10000,
+      "public_split_name": "validation",
+      "repeated_inputs": 2831,
+      "source_split_name": "val",
+      "strategy_counts": {
+        "base_conversion:decomposition": 574,
+        "base_conversion:repeated-division": 524,
+        "base_conversion:via-binary": 578,
+        "grid_pathfinding:alternating": 518,
+        "grid_pathfinding:down-first": 548,
+        "grid_pathfinding:right-first": 562,
+        "linear_equation_solving:divide-then-subtract": 594,
+        "linear_equation_solving:inverse-ops": 535,
+        "linear_equation_solving:subtract-then-divide": 545,
+        "list_summation:left-to-right": 541,
+        "list_summation:pairwise": 553,
+        "list_summation:right-to-left": 580,
+        "multidigit_addition:left-to-right-partials": 527,
+        "multidigit_addition:right-to-left-carry": 555,
+        "multidigit_addition:rounding-decomposition": 551,
+        "sorting_algorithms:bubble-sort": 350,
+        "sorting_algorithms:heap-sort": 325,
+        "sorting_algorithms:insertion-sort": 341,
+        "sorting_algorithms:merge-sort": 368,
+        "sorting_algorithms:selection-sort": 331
+      },
+      "task_counts": {
+        "base_conversion": 1676,
+        "grid_pathfinding": 1628,
+        "linear_equation_solving": 1674,
+        "list_summation": 1674,
+        "multidigit_addition": 1633,
+        "sorting_algorithms": 1715
+      },
+      "unique_inputs": 7169,
+      "unique_rows": 8318
+    }
+  },
+  "strategy_ids": [
+    "base_conversion:decomposition",
+    "base_conversion:repeated-division",
+    "base_conversion:via-binary",
+    "grid_pathfinding:alternating",
+    "grid_pathfinding:down-first",
+    "grid_pathfinding:right-first",
+    "linear_equation_solving:divide-then-subtract",
+    "linear_equation_solving:inverse-ops",
+    "linear_equation_solving:subtract-then-divide",
+    "list_summation:left-to-right",
+    "list_summation:pairwise",
+    "list_summation:right-to-left",
+    "multidigit_addition:left-to-right-partials",
+    "multidigit_addition:right-to-left-carry",
+    "multidigit_addition:rounding-decomposition",
+    "sorting_algorithms:bubble-sort",
+    "sorting_algorithms:heap-sort",
+    "sorting_algorithms:insertion-sort",
+    "sorting_algorithms:merge-sort",
+    "sorting_algorithms:selection-sort"
+  ]
+}