nutshells3's picture
Release LiveGraph frozen benchmark
438f687 verified
|
Raw
History Blame Contribute Delete
2.8 kB
---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- reasoning
- graph-reasoning
- synthetic
- verifier
- paired-evaluation
pretty_name: LiveGraph Matched Graph-vs-Linear Benchmark
---
# LiveGraph Matched Graph-vs-Linear Benchmark
This release contains the frozen synthetic datasets used by *LiveGraph: Typed
Graphs versus Linear Traces*. It supports a
controlled comparison between a typed-graph rendering and a chronological
linear-trace rendering of the same Boolean relational-propagation records.
## Main matched benchmark
- `data/semantic_train.jsonl`: 480 semantic training records, depths 1--10.
- `data/semantic_heldout.jsonl`: 3,200 clean held-out records, depths 1--32.
- `data/semantic_eval_variants.jsonl`: 19,200 semantic evaluation variants.
- `data/sft_{typed_graph,linear_trace}.jsonl`: paired SFT rows.
- `data/eval_{typed_graph,linear_trace}.jsonl`: paired evaluation rows.
- `manifest.json`: frozen counts, construction invariants, and file hashes.
The variants are `clean`, `edge_order`, `node_rename`,
`extra_distractors`, `one_path_edge_flip`, and `edge_delete`. The generator
balances labels within depth and holds the semantic record and full-chat token
count fixed across the two registered renderings. Ordinary visible trailing
`PAD` markers are part of the graph rendering package; this is not a pure
graph-topology intervention.
## Additional frozen subsets
- `data/grpo_v1/`: exact-verifier GRPO train, ID, and OOD sets.
- `data/native_robustness_v1/`: native-thinking robustness subset.
- `data/edge_delete_unknown_v1/`: deletion-only UNKNOWN evaluation.
- `data/subsets_v3/`: clean and diagnostic subsets.
- `data/relassum_v3/`: auxiliary relation/assumption evaluation artifacts.
## Integrity
Verify the frozen main files from this directory:
```bash
sha256sum -c SHA256SUMS
```
For the original main files, the frozen hashes are also recorded in
`manifest.json` and `SHA256SUMS`. Derived subset directories carry their own
manifests.
The byte-frozen top-level manifest retains two original absolute tokenizer and
chat-template provenance strings so that its paper-reported SHA-256 remains
verifiable. They are inert local paths; use the public base model files rather
than those locations.
## Scope and limitations
The data are synthetic. `supports` preserves a Boolean value and `refutes`
negates it along one queried path. These tasks do not establish real-world
reasoning quality, theorem validity, source-claim truth, or end-to-end OAE
assurance. Auxiliary relation/assumption labels are not human gold.
## License and citation
The released dataset is provided under Apache License 2.0. Cite the LiveGraph
paper and this dataset release. A final arXiv identifier can be added to this
card after submission.