zimplex/dllm-random-parents-llada2-finecode-edge-matched-b64-k8-seed20260924-v1
DART LLaDA2 FineCode edge-matched random dependency graph This is the sealed random-graph control for LLaDA 2 Mini FineCode experiments. It pairs only with zimplex/dllm-finecode-dagcover-full-v1, whose manifest SHA-256 is 45e250f9f4a5302ef1bb04c16577eb8b6d85996368fe1dc359b86465cdb03176. For every independent document-local B64 mutable unit, the generator matches the exact effective post-K8 edge count of corrected attention pred-debias5, samples a deterministic random topological… See the full description on the dataset page: https://huggingface.co/datasets/zimplex/dllm-random-parents-llada2-finecode-edge-matched-b64-k8-seed20260924-v1.
DART LLaDA2 FineCode edge-matched random dependency graph
This is the sealed random-graph control for LLaDA 2 Mini FineCode experiments. It pairs only with `zimplex/dllm-finecode-dagcover-full-v1`, whose manifest SHA-256 is 45e250f9f4a5302ef1bb04c16577eb8b6d85996368fe1dc359b86465cdb03176.
For every independent document-local B64 mutable unit, the generator matches the exact effective post-K8 edge count of corrected attention pred-debias5, samples a deterministic random topological order, and draws an acyclic random graph with maximum indegree eight. It does not preserve degree, distance, root, depth, or sequence-causal statistics. The global seed is 20260924.
Integrity and use
manifest.jsonSHA-256:375d5c087fe7a33bcaa3270d056184b146f58f0464b7504dddb17556cf7d9a40- Records: 1,713,961
- Edges: 1,548,477,379
- Sequence length: 2,048
- Training block size: 64
- Maximum parents: 8
The repository contains no raw text. keys.npy maps exact packed examples to parents.npy and weights.npy. The graph is not interchangeable across tokenizers, packing layouts, corpus variants, or dataset manifests.
Download the complete repository to regular files on shared storage, then run sha256sum -c SHA256SUMS. Retain manifest.json, COMPLETE.json, and PAIRING.json; they are the authoritative identity and source binding. A Hugging Face cache containing symlinks is not a valid direct training path.
This is a memory-mapped runtime artifact, not a streaming dataset. The public availability of this derived dependency annotation does not assert a separate permissive license.