SLWM-X450M-Memory / README.md
diogoantunes's picture
Document examples and validation; align MIT copyright contacts
aff4089 verified
|
Raw History Blame Contribute Delete
8.47 kB
---
license: mit
language:
- en
library_name: jax
tags:
- world-model
- multimodal
- working-memory
- embodied-control
- research
- study-artifacts
- npz
---
# SLWM-X450M — Study artifacts (Wave 6 memory reference)
> **Study artifacts for reference — unpublished manuscript.** These checkpoints accompany the manuscript draft v0.3, which has not yet been published. They are shared for inspection and reproducibility of the study, not as a formal model release.
The full 453M-parameter, Epps-Pulley SIGReg reference from the single-factor ablation study. It preserves key shielding, the sensory bypass, and decoupled readout heads.
Research by **Diogo Antunes**, at [valuealigned.ai](https://valuealigned.ai) and [keplen.ai](https://keplen.ai). Shared as study artifacts in the **valueai-org** organization.
This is an original, unmodified checkpoint from the SLWM research archive. It uses a **custom NumPy NPZ format** and the SLWM JAX runtime; it is not a Transformers `AutoModel` checkpoint or a chat model.
## Study artifact reference
| Property | Value |
|---|---|
| Original run | `x1b-w6_slwm_450m_full_epps_sigreg-s2001` |
| Experimental wave | 6 |
| Training seed | 2001 |
| Strict total parameters | 453,199,938 |
| Training steps | 2,600 |
| Batch size | 12 sequences |
| Training sequence length | 32 ticks |
| Latent width / layers / slots per tick | 1152 / 18 / 24 |
| Runtime/source reference | [`cbb2b38`](https://github.com/ValueAlignedAI/slwm/tree/cbb2b38794359adfefe07b19e9b7caa113b25b47) |
Seed 2001 is provided as a fixed reference, not selected by ranking test scores. The unpublished manuscript reports three-seed cohorts; this repository contains one checkpoint and its own recorded metrics.
This is the Wave 6 full model with end-effector-only proprioception. It is useful as a reference for high-cardinality dynamic-visual overwrite memory and the global-pooling spatial bottleneck. It is not the Wave 7 spatial refinement.
## Files and integrity
| File | Bytes | SHA-256 |
|---|---:|---|
| `weights_bf16.npz` | 719,568,210 | `fe63862af1a8d3440b0a8b1e580b4961a179c9a65b7da0957bf25c434dab45a8` |
| `portable_int8.npz` | 415,554,792 | `36d94dfe6db11a0d06b73f816bfd5a32c17b29865ab58ca4b778739dfb6a0e89` |
- `weights_bf16.npz`: the project's mixed BF16/FP32 export. Matrix BF16 values are stored as uint16 bit patterns, with a dtype manifest.
- `portable_int8.npz`: the project's symmetric per-channel INT8 export, with scales and a quantization manifest.
- `config.json`, `receipt.json`, `eval_metrics.json`, and `train.jsonl`: original run artifacts, preserved byte for byte.
- `artifact_manifest.json` and `SHA256SUMS`: source mapping, artifact hashes, and export-validation results.
- `load_checkpoint.py`: a small, pickle-free NumPy loader for both formats.
All original files were checked against their recorded SHA-256 values before upload. Both weight exports were decoded on CPU, checked for finite values, and counted against the strict total and component counts. These are packaging/integrity checks, not new benchmark evaluations.
The provided loader expands weights to float32. **INT8 file size is not the runtime memory requirement, and this loader does not perform INT8 matrix computation.** Model weights, activations, and planning buffers are separate from the paper's bounded streaming-state figures.
## Download and load
Install `numpy` and `huggingface_hub` in your Python environment, then:
```python
import json
import sys
from pathlib import Path
from huggingface_hub import snapshot_download
folder = Path(snapshot_download(
repo_id="valueai-org/SLWM-X450M-Memory",
allow_patterns=["portable_int8.npz", "config.json", "receipt.json", "load_checkpoint.py"],
))
sys.path.insert(0, str(folder))
from load_checkpoint import load_checkpoint
receipt = json.loads((folder / "receipt.json").read_text())
params = load_checkpoint(
folder / "portable_int8.npz",
expected_sha256=receipt["portable_int8_sha256"],
)
assert sum(value.size for value in params.values()) == receipt["parameter_counts"]["strict_total"]
```
For the BF16 export, download `weights_bf16.npz` instead and use `receipt["weights_bf16_sha256"]` for verification. `params` is a dictionary of NumPy float32 arrays. JAX execution requires JAX and the appropriate modules from the [pinned research source](https://github.com/ValueAlignedAI/slwm/tree/cbb2b38794359adfefe07b19e9b7caa113b25b47); follow that source for input preparation, modality masks, state updates, and planning.
Use `models/x1b_wave6_core.py`, including `Wave6Interventions` and `forward_world_model_wave6`. Preserve the `interventions` in `config.json`. The matching evaluation entry points are in `evals/x1b_eval_wave6.py`.
## Run an example and validation checks
Use the [runnable examples and validation-set guide](https://github.com/ValueAlignedAI/slwm/blob/main/docs/research/x1b_study_artifacts_and_validation.md) from the SLWM source repository. It explains the development and held-out data splits, how to verify local data-bank checksums, and which recorded inputs are not included with these checkpoints.
From a current checkout of [the source repository](https://github.com/ValueAlignedAI/slwm), after installing the guide's Python dependencies:
```sh
python scripts/eval/x1b_study_examples.py \
--mode example --repo-id valueai-org/SLWM-X450M-Memory --format int8 \
--output /tmp/slwm-x450m-memory-example.json
```
This is a generated, deterministic functional example using the selected trained checkpoint; it is not a benchmark score. The guide provides BF16 commands for the original evaluation functions and the exact bank paths and split names for validation.
## Recorded results for this checkpoint
These values come directly from this run's `receipt.json`. They are **single-seed results**, not the manuscript's cohort mean ± standard deviation. The upload does not claim that the quantized export has received an additional independent evaluation.
| Recorded metric | Seed 2001 |
|---|---:|
| Dynamic-visual overwrite recall at T = 1,024 | 100.0000 % |
| Unseen-layout MPPI success | 1.5625 % |
| Unseen camera-corruption AUROC | 50.3200 % |
| Held-out-scene xArm6 four-step error | 33.2159 mm |
The original `eval_metrics.json` provides the wider evaluation record, including failures. Compare architectures within the same wave and observation/training setting.
## Scope and limitations
The strong associative-recall result coexists with weak unprivileged visual planning and near-chance unseen camera-fault detection in this wave. Do not substitute this checkpoint for SLWM-v2 when evaluating the later spatial-grid planner. Physical xArm6 results remain offline recorded-trajectory evaluations.
The checkpoint is intended for research and reproducibility. The upload contains no raw training recordings, datasets, optimizer state, or private server access configuration. The pinned source commit is a public runtime/study reference, not a separately verified training-code commit: the archived receipts do not record the training Git SHA.
## Sources and licensing
- [Unpublished manuscript draft v0.3 (PDF)](https://diogoantunes.ai/media/publications/slwm_x1b_paper_v03.pdf)
- [Expanded article, figures, and experiment context](https://diogoantunes.ai/publications/signal-latent-world-models)
- [SLWM source repository](https://github.com/ValueAlignedAI/slwm/tree/cbb2b38794359adfefe07b19e9b7caa113b25b47)
- [Original experiment records](https://github.com/ValueAlignedAI/slwm/tree/cbb2b38794359adfefe07b19e9b7caa113b25b47/experiments/multimodal/x1b/spatiotemporal-jepa-1b-20261111/runs_wave6/x1b-w6_slwm_450m_full_epps_sigreg-s2001)
These study artifacts, including the BF16 and INT8 weights, use the **MIT license**, as specified by the author. The complete `LICENSE` and `SOURCE_LICENSE` files match the [corrected GitHub project license](https://github.com/ValueAlignedAI/slwm/blob/d1abfc031d9d0c54f511e9a4eb324dfc0b57a08d/LICENSE). The copyright owners are **Diogo Antunes** ([contact@diogoantunes.ai](mailto:contact@diogoantunes.ai)), **ValueAI** ([contact@valuealigned.ai](mailto:contact@valuealigned.ai)), and **Keplen** ([contact@keplen.ai](mailto:contact@keplen.ai)).
## Citation
Diogo Antunes. *Signal-Latent World Models at 1.008B Scale: Separating Fast Sensory Grounding from Editable Working Memory Under O(1) Streaming Control.* Unpublished manuscript, draft v0.3, dated November 14, 2026.