schema_version string | code_repository string | model_repository string | dataset_repository string | collection string | artifacts list | notes list |
|---|---|---|---|---|---|---|
1.0 | https://github.com/HLR/CLAMP | https://huggingface.co/SueMintony/CLAMP-HMM-Checkpoints | https://huggingface.co/datasets/SueMintony/CLAMP-Sampled-Continuations-and-Demos | https://huggingface.co/collections/SueMintony/clamp-hmm-checkpoints-sampled-continuations-and-demos-6a876ddfa7602097020a565c | [
{
"family": "Qwen3.5-9B",
"domain": "behavior",
"variant": "combined-v2",
"kind": "trained_hmm",
"remote": "trained/qwen3.5-9b/behavior/combined-v2/hmm-h128-lr1e-3-ep50/checkpoint.eqx",
"bytes": 508260236,
"sha256": "d07e382e4eb192f969540e2ee9c6b1b7266f1c45dabb3c7acc8eb26fc67420a4"
},
... | [
"Qwen3.5 NPZ files contain sampled continuation token arrays plus lengths and padding metadata; original prompts and scene assets are not included in this public release.",
"InternVL3.5 samples.jsonl contains 2,926 sampled-continuation records from IV35-H2-S5.",
"The Qwen3-VL artifact is a compatibility vocabul... |
CLAMP Sampled Continuations and VLABench Demos
This dataset accompanies HLR/CLAMP, the Constrained Language-Action Model Planner.
It contains 20 closed-loop VLABench case studies: 10 successful and 10 unsuccessful executions. Each case includes the input image and mask, instruction, prompt and entity metadata, CLAMP and ground-truth plans, evaluation output, execution video, and a frame manifest.
Layout
demo_dataset/
├── manifest.json
├── README.md
├── success/<task>/
└── failure/<task>/
The success and failure split is retained for transparent qualitative analysis; it is not a training split. See demo_dataset/README.md and demo_dataset/manifest.json for the full schema and per-example provenance.
Related release
- Code: HLR/CLAMP
- HMM checkpoints: SueMintony/CLAMP-HMM-Checkpoints
- Release collection: CLAMP — HMM Checkpoints, Sampled Continuations, and Demos
Sampled continuations and checkpoint pairing
training-artifacts/release_catalog.json provides the exact Model-repository path, Dataset-repository path, byte size, SHA-256 digest, family, domain, and variant for every release artifact.
sampled-continuations/qwen3.5-9b/pairs with the 10 Qwen3.5 HMM checkpoints in the Model repository.sampled-continuations/internvl3.5-8b/iv35-h2-s5/pairs with the IV35-H2-S5 HMM release.- The Qwen3-VL-compatible HMM is a model-only vocabulary-remap artifact, not a trained model; no local sampled-continuation file exists for it.
Qwen3.5 .npz files contain continuation token arrays, lengths, maximum-length, and padding metadata. Original prompts, scene assets, and benchmark data are intentionally excluded.
- Downloads last month
- 33