README.md

---
license: apache-2.0
tags:
  - hmm
  - embodied-agents
  - ctrlg
  - constrained-decoding
---
# CLAMP HMM Checkpoints

Neural-HMM priors trained on Llama-3.1 continuation samples for two EAI
datasets (BEHAVIOR and VirtualHome). Used by `eai_ctrlg` as the semantic
prior for constrained decoding with γ+β DFAs.

## Files
- `behavior/hmm-h128-lr0.01/checkpoint.eqx` — hidden=128, lr=0.01
- `virtualhome/hmm-h128-lr0.01/checkpoint.eqx` — hidden=128, lr=0.01

## Loading
```python
import equinox as eqx
from ctrlg.hmm.model import ConditionalHMM
model = eqx.tree_deserialise_leaves(
    "behavior/hmm-h128-lr0.01/checkpoint.eqx", ConditionalHMM(...))
```

Training pipeline: see `eai_train/cond_hmm/` in the companion repo.

## Experimental checkpoint: IV35-H2-S5 (InternVL3.5)

`experimental/iv35-h2-s5-internvl3.5-8b/hmm_formal/checkpoint.eqx` is an experimental neural-HMM prior trained from frozen `OpenGVLab/InternVL3_5-8B` continuations. It is **not** an InternVL base-model checkpoint and does not replace the two stable BEHAVIOR and VirtualHome checkpoints above.

- Training status: PASS; independent 50-epoch formal run
- Training / validation sequences: 13,170 / 1,460
- Best validation loss: 2.235296
- SHA-256: `3c5f90c6ed76879ba465f7dd009e62a9c87a62905e8f77b8dfffe428527d2512`
- Reproducibility metadata: `experimental/iv35-h2-s5-internvl3.5-8b/hmm_formal/release_metadata.json`

The checkpoint passed its recorded training and reload validation gates. No downstream task-gain claim is made for this experimental release.

## Related data and code

- Code: [HLR/CLAMP](https://github.com/HLR/CLAMP)
- VLABench demonstrations: [SueMintony/CLAMP-Sampled-Continuations-and-Demos](https://huggingface.co/datasets/SueMintony/CLAMP-Sampled-Continuations-and-Demos)
- Release collection: [CLAMP — HMM Checkpoints, Sampled Continuations, and Demos](https://huggingface.co/collections/SueMintony/clamp-hmm-checkpoints-sampled-continuations-and-demos-6a876ddfa7602097020a565c)

<!-- CLAMP_TRAINING_ARTIFACTS_START -->
## Published training artifacts

The full, machine-readable pairing of every checkpoint and sampled-continuation artifact is in [`release_catalog.json`](release_catalog.json).

- **Qwen3.5-9B:** 10 archived HMM checkpoints under [`trained/qwen3.5-9b/`](trained/qwen3.5-9b) — combined BEHAVIOR and VirtualHome variants plus four task-specific variants per domain. Their continuation files are in the [Dataset repository](https://huggingface.co/datasets/SueMintony/CLAMP-Sampled-Continuations-and-Demos/tree/main/sampled-continuations/qwen3.5-9b).
- **InternVL3.5-8B:** IV35-H2-S5 formal EQX checkpoint, portable NPZ export, and [2,926 sampled records](https://huggingface.co/datasets/SueMintony/CLAMP-Sampled-Continuations-and-Demos/tree/main/sampled-continuations/internvl3.5-8b/iv35-h2-s5).
- **Qwen3-VL-8B-Instruct:** [compatibility-remapped HMM](compatibility/qwen3-vl-8b-instruct/behavior/hmm-h128-remapped/checkpoint.eqx). It is a vocabulary remap of a historical HMM, not a retrained model, and has no local sampled-continuation artifact.

The public continuation release contains token arrays / sampled outputs and length metadata only; it excludes original prompts, scene assets, and benchmark data.
<!-- CLAMP_TRAINING_ARTIFACTS_END -->