redwoodresearch/sdf-twohop-data

sdf-twohop: the data behind the paper

Training data, per-item evaluations, run records and figures of Small LLMs perform no-CoT two-hop reasoning on synthetic facts with real bridges (code: https://github.com/redwoodresearch/sdf-twohop). 3,614 files, 5.2 GiB, laid out exactly as the code repository expects, so downloading into a clone reproduces the working tree every script reads:

git clone https://github.com/redwoodresearch/sdf-twohop && cd sdf-twohop && uv sync
uv run scripts/fetch_release.py --figures    # grids and figures, ~100 MB: fancy_figures.py redraws them
uv run scripts/fetch_release.py --evals      # per-item evaluations, 3.1 GB: bridge_compare.py, cell_grid.py
uv run scripts/fetch_release.py --runs       # run records, 100 MB: eval_checkpoint.py on the Tinker checkpoints
uv run scripts/fetch_release.py --training   # corpora and C4 sample, 2.1 GB: train_sdf.py --order-from <run>

The matching LoRA adapters are under redwoodresearch/sdf-twohop-lora-qwen3.6-35b-a3b and redwoodresearch/sdf-twohop-lora-qwen3-8b (final checkpoint of every run, one subfolder per run; every checkpoint is also public on Tinker under the paths in runs/<run>/checkpoints.jsonl).

Layout

• data/generated/*.json(l): the frozen entities and evaluation items (also in the code repository).
• data/generated/corpus/<corpus>/: the training data. docs.jsonl of a document corpus holds every document that passed the filters and the audit, with its verdict fields; docs.jsonl of a qa_* corpus holds the Q/A rows. A run saw the first 375 per fact of a seeded shuffle; runs/<run>/doc_order.json lists exactly which, in order.
• data/upstream/c4_50000.jsonl: the web-text sample mixed into the *_c4 runs (allenai/c4, ODC-BY).
• runs/<run>/: training config, document order, training log, and the Tinker checkpoint paths.
• results/evals/<run>/<checkpoint>/: items.jsonl (every prompt, completion and per-candidate teacher-forced NLL of every evaluation item), beliefs.jsonl (the single-fact recall probes) and summary.json; base/ is the untrained model on the same items.
• results/figures/*_grid.json: the per-checkpoint statistics the figures draw; results/fancyfigures/: the figures (PNG, PDF) and their tables.
• manifest.json: sha256, size, the runs and the paper figures behind every file.

Paper figures

Figure	File	Grid	Runs
2	results/fancyfigures/png/020_bridge_exp_sft_lr1.6e-4.png	bridge_grid.json	5 runs
3	results/fancyfigures/png/030_bridge_exp_sdf_lr5e-4.png	dbridge_grid.json	4 runs
4	results/fancyfigures/png/040_hops_exp_final.png	qa_cellgrid.json, d375_cellgrid.json	3 runs
A1	results/fancyfigures/png/A10_bridge_exp_sft_other_lr1.6e-4.png	bridge_grid.json	5 runs
A2	results/fancyfigures/png/A20_bridge_exp_sdf_other_lr5e-4.png	dbridge_grid.json	4 runs
A3	results/fancyfigures/png/A30_hops_exp_final_other.png	qa_cellgrid.json, d375_cellgrid.json	3 runs
B1	results/fancyfigures/png/B10_bridge_exp_sft_lr4.7e-4_8b.png	q8bridge_grid.json	5 runs
B2	results/fancyfigures/png/B20_bridge_exp_sdf_lr4.7e-4_8b.png	q8dbridge_grid.json	4 runs
B3	results/fancyfigures/png/B30_hops_exp_final_8b.png	q8qa_cellgrid.json, q8_cellgrid.json	3 runs
C1	results/fancyfigures/png/C10_bridge_exp_sft_lr_ablation.png	bridge_grid.json	16 runs
C2	results/fancyfigures/png/C20_bridge_exp_sdf_lr_ablation.png	dbridge_grid.json	13 runs
C3	results/fancyfigures/png/C30_hops_exp_sft_lr_ablation.png	qa_cellgrid.json	4 runs
C4	results/fancyfigures/png/C40_hops_exp_sdf_lr_ablation.png	d375_cellgrid.json	4 runs
C5	results/fancyfigures/png/C50_hops_exp_sdf_mixed_lr_ablation.png	d375_cellgrid.json	4 runs

Runs

Peak learning rates are the Tinker cookbook's LoRA rate for each model and half-decade steps below it. "Figures" lists where a run appears; runs without one are the learning-rate sweep described in Appendix C.

Run	Model	Bridge entity	Recipe	Peak LR	Figures
qa_lr5e-4	Qwen3.6-35B-A3B	real country	SFT	5e-4	C1, C3
qa_lr1.6e-4	Qwen3.6-35B-A3B	real country	SFT	1.6e-4	2, 4, A1, A3, C1, C3
qa_lr5e-5	Qwen3.6-35B-A3B	real country	SFT	5e-5	C1, C3
qa_lr1.6e-5	Qwen3.6-35B-A3B	real country	SFT	1.6e-5	C1, C3
d375_lr5e-4	Qwen3.6-35B-A3B	real country	SDF	5e-4	3, 4, A2, A3, C2, C4
d375_lr5e-4_c4	Qwen3.6-35B-A3B	real country	SDF + C4	5e-4	4, A3, C5
d375_lr1.6e-4	Qwen3.6-35B-A3B	real country	SDF	1.6e-4	C2, C4
d375_lr1.6e-4_c4	Qwen3.6-35B-A3B	real country	SDF + C4	1.6e-4	C5
d375_lr5e-5	Qwen3.6-35B-A3B	real country	SDF	5e-5	C2, C4
d375_lr5e-5_c4	Qwen3.6-35B-A3B	real country	SDF + C4	5e-5	C5
d375_lr1.6e-5	Qwen3.6-35B-A3B	real country	SDF	1.6e-5	C2, C4
d375_lr1.6e-5_c4	Qwen3.6-35B-A3B	real country	SDF + C4	1.6e-5	C5
qafamous_lr5e-4	Qwen3.6-35B-A3B	famous city	SFT	5e-4	C1
qafamous_lr1.6e-4	Qwen3.6-35B-A3B	famous city	SFT	1.6e-4	2, A1, C1
qafamous_lr5e-5	Qwen3.6-35B-A3B	famous city	SFT	5e-5	C1
dfamous_lr5e-4	Qwen3.6-35B-A3B	famous city	SDF	5e-4	3, A2, C2
dfamous_lr1.6e-4	Qwen3.6-35B-A3B	famous city	SDF	1.6e-4	C2
dfamous_lr5e-5	Qwen3.6-35B-A3B	famous city	SDF	5e-5	C2
qaobscure_lr5e-4	Qwen3.6-35B-A3B	obscure city	SFT	5e-4	C1
qaobscure_lr1.6e-4	Qwen3.6-35B-A3B	obscure city	SFT	1.6e-4	2, A1, C1
qaobscure_lr5e-5	Qwen3.6-35B-A3B	obscure city	SFT	5e-5	C1
dobscure_lr5e-4	Qwen3.6-35B-A3B	obscure city	SDF	5e-4	3, A2, C2
dobscure_lr1.6e-4	Qwen3.6-35B-A3B	obscure city	SDF	1.6e-4	C2
dobscure_lr5e-5	Qwen3.6-35B-A3B	obscure city	SDF	5e-5	C2
qacity_lr5e-4	Qwen3.6-35B-A3B	fabricated city	SFT	5e-4	C1
qacity_lr1.6e-4	Qwen3.6-35B-A3B	fabricated city	SFT	1.6e-4	2, A1, C1
qacity_lr5e-5	Qwen3.6-35B-A3B	fabricated city	SFT	5e-5	C1
dcity_lr5e-4	Qwen3.6-35B-A3B	fabricated city	SDF	5e-4	3, A2, C2
dcity_lr1.6e-4	Qwen3.6-35B-A3B	fabricated city	SDF	1.6e-4	C2
dcity_lr5e-5	Qwen3.6-35B-A3B	fabricated city	SDF	5e-5	C2
qaword_lr5e-4	Qwen3.6-35B-A3B	Spouses city	SFT	5e-4	C1
qaword_lr1.6e-4	Qwen3.6-35B-A3B	Spouses city	SFT	1.6e-4	2, A1, C1
qaword_lr5e-5	Qwen3.6-35B-A3B	Spouses city	SFT	5e-5	C1
q8qa_lr4.7e-4	Qwen3-8B	real country	SFT	4.7e-4	B1, B3
q8qa_lr1.5e-4	Qwen3-8B	real country	SFT	1.5e-4	learning-rate sweep (Appendix C text only)
q8qa_lr4.7e-5	Qwen3-8B	real country	SFT	4.7e-5	learning-rate sweep (Appendix C text only)
q8qa_lr1.5e-5	Qwen3-8B	real country	SFT	1.5e-5	learning-rate sweep (Appendix C text only)
q8_lr4.7e-4	Qwen3-8B	real country	SDF	4.7e-4	B2, B3
q8_lr4.7e-4_c4	Qwen3-8B	real country	SDF + C4	4.7e-4	B3
q8_lr1.5e-4	Qwen3-8B	real country	SDF	1.5e-4	learning-rate sweep (Appendix C text only)
q8_lr1.5e-4_c4	Qwen3-8B	real country	SDF + C4	1.5e-4	learning-rate sweep (Appendix C text only)
q8_lr4.7e-5	Qwen3-8B	real country	SDF	4.7e-5	learning-rate sweep (Appendix C text only)
q8_lr4.7e-5_c4	Qwen3-8B	real country	SDF + C4	4.7e-5	learning-rate sweep (Appendix C text only)
q8_lr1.5e-5	Qwen3-8B	real country	SDF	1.5e-5	learning-rate sweep (Appendix C text only)
q8_lr1.5e-5_c4	Qwen3-8B	real country	SDF + C4	1.5e-5	learning-rate sweep (Appendix C text only)
q8famous_lr4.7e-4	Qwen3-8B	famous city	SFT	4.7e-4	B1
q8famous_lr1.5e-4	Qwen3-8B	famous city	SFT	1.5e-4	learning-rate sweep (Appendix C text only)
q8famous_lr4.7e-5	Qwen3-8B	famous city	SFT	4.7e-5	learning-rate sweep (Appendix C text only)
q8dfamous_lr4.7e-4	Qwen3-8B	famous city	SDF	4.7e-4	B2
q8dfamous_lr1.5e-4	Qwen3-8B	famous city	SDF	1.5e-4	learning-rate sweep (Appendix C text only)
q8dfamous_lr4.7e-5	Qwen3-8B	famous city	SDF	4.7e-5	learning-rate sweep (Appendix C text only)
q8obscure_lr4.7e-4	Qwen3-8B	obscure city	SFT	4.7e-4	B1
q8obscure_lr1.5e-4	Qwen3-8B	obscure city	SFT	1.5e-4	learning-rate sweep (Appendix C text only)
q8obscure_lr4.7e-5	Qwen3-8B	obscure city	SFT	4.7e-5	learning-rate sweep (Appendix C text only)
q8dobscure_lr4.7e-4	Qwen3-8B	obscure city	SDF	4.7e-4	B2
q8dobscure_lr1.5e-4	Qwen3-8B	obscure city	SDF	1.5e-4	learning-rate sweep (Appendix C text only)
q8dobscure_lr4.7e-5	Qwen3-8B	obscure city	SDF	4.7e-5	learning-rate sweep (Appendix C text only)
q8city_lr4.7e-4	Qwen3-8B	fabricated city	SFT	4.7e-4	B1
q8city_lr1.5e-4	Qwen3-8B	fabricated city	SFT	1.5e-4	learning-rate sweep (Appendix C text only)
q8city_lr4.7e-5	Qwen3-8B	fabricated city	SFT	4.7e-5	learning-rate sweep (Appendix C text only)
q8dcity_lr4.7e-4	Qwen3-8B	fabricated city	SDF	4.7e-4	B2
q8dcity_lr1.5e-4	Qwen3-8B	fabricated city	SDF	1.5e-4	learning-rate sweep (Appendix C text only)
q8dcity_lr4.7e-5	Qwen3-8B	fabricated city	SDF	4.7e-5	learning-rate sweep (Appendix C text only)
q8word_lr4.7e-4	Qwen3-8B	Spouses city	SFT	4.7e-4	B1
q8word_lr1.5e-4	Qwen3-8B	Spouses city	SFT	1.5e-4	learning-rate sweep (Appendix C text only)
q8word_lr4.7e-5	Qwen3-8B	Spouses city	SFT	4.7e-5	learning-rate sweep (Appendix C text only)

Scope and provenance

This is Tier A of the project's artefacts: everything the paper's figures and tables are computed from. The pilots, the longer fabricated-city runs and the other analyses described in the code repository's README ("Beyond the paper") are not included here.

Documents and Q/A rows are AI-generated synthetic text (Claude Opus 5.5, Sonnet 5.5 and Haiku 4.5 through the Believe It or Not pipeline) about fictional universities and registries, real countries and real cities; they contain no personal data. Sources: SOCRATES (CC-BY-4.0) for the real two-hop items; Wikidata (CC0) for countries and cities; allenai/c4 (ODC-BY) for the web-text sample; Qwen3.6-35B-A3B and Qwen3-8B (Apache-2.0) as base models.