akumch/graph-reasoning-foundational-curriculum

Foundational Structural Curriculum — v1.1

TEMPORARY MIRROR. This repository is a convenience copy for browsing and exploration, published so a reader can look at the data without S3/DVC credentials. It is not the canonical dataset and is not guaranteed to stay in sync. The source of truth is the DVC-tracked corpus in the graph-reasoning-llm repo (pin by the git SHA of dvc.lock, never "latest"). The usual byte-for-byte verification against dvc.lock was skipped for this upload. Treat these files as a snapshot that may be removed or replaced.

A procedurally generated table-reasoning curriculum plus sealed evaluation batteries, built for training a Llama-3.2-1B model that reasons over in-context structured tables.

What's here

Also included: splits.json, curriculum/build_manifest.json, battery/battery_manifest.json (per-battery n + sha256), and DATACARD.md (the full committed data card with per-task / per-source / per-split counts).

Record schema

Each line is one JSON object. Core fields shared by every record:

• —question — the natural-language query.
• —gold — the reference answer (a | -joined string for multi-value answers).
• —columns — list of column names.
• —rows — the table body as a list of rows.
• —cells — the table as [row_idx, col_idx]-addressable cells.
• —qtype — task type (extract_column, argmax, chained_lookup, ...).
• —source — scm (synthetic structural-causal-model world) or mmtu (real MMTU table grid).
• —n_rows, n_cols, target_col_idx, target_col_name, target_row_idx, id, source_id, split.

Battery records add task-specific fields (filter_type, hop_count, key_col_idx, key_col_name, key_value, n_matches).

Records are heterogeneous by design: each qtype carries its own task-specific fields, so curriculum/train.jsonl holds 14 distinct key sets across 35 columns. The card declares the full column set per config, and the loader fills the fields a given record does not use with null. Without that declaration the viewer infers the schema from the first chunk of the file and fails on the first later chunk that introduces a new column.

Load

from datasets import load_dataset

train = load_dataset("akumch/graph-reasoning-foundational-curriculum", "curriculum", split="train")
rowaddr = load_dataset("akumch/graph-reasoning-foundational-curriculum", "eval_rowaddr", split="train")
print(train[0]["question"], "->", train[0]["gold"])

Or just browse the Data Studio viewer tab above — no download needed.

Provenance

• —Curriculum version: foundational-curriculum-v1.1
• —Eval version: foundational-eval-v1
• —Base model: meta-llama/Llama-3.2-1B
• —World bank: 1,000 SCM worlds, bank seed 20260711
• —Sources: synthetic SCM worlds + real MMTU table grids
• —Snapshot taken from graph-reasoning-llm at commit 7c43edd (dvc.lock last changed at 7882746).

License

Released under CC-BY-4.0. The synthetic SCM portion is original. The mmtu portion re-hosts tables drawn from the MMTU benchmark; if a different license is needed for that portion, open an issue on the dataset repo and it will be added.