juiceb0xc0de/mmb3-slt-6xcam-results

Mamba-3 187M SISO — Training Observation Bundles

Raw measurement artifacts from training state-spaces/mamba3-siso-187m on the ordered GSM8K training split. Six independent cameras observe the same optimizer update at fixed, predeclared intervals.

Artifact archive, not a loadable dataset. These are deterministic NPZ bundles with manifests and checksums. The dataset viewer is disabled because there is no tabular schema to render. No analysis code or derived results are included here.

Current contents

Directory	Status	Training covered	Observation bundles	Size
observed-runs/20260810T091311Z/	complete run 1	2 × 7,473 rows = 14,946 updates	148	104.53 GB
observed-runs/20260810T115113Z/	retained partial replication (run2)	stopped at update 2,600	25, through update 2,500	18.52 GB

Run 1 is the primary corpus. It observes every 100th epoch-local update from 100 through 7,400 in each of two epochs. Its run_summary.json records all 148 observation points. W&B run: hu1wjq02.

Run 2 was launched as a warm-snapshot replication with 25 warmup rows. It did not complete the declared two-epoch schedule, so it is explicitly retained as a partial run rather than presented as a second full corpus. W&B run: s5mwzyjy.

Shared provenance

Model	state-spaces/mamba3-siso-187m
Revision	6792c27c00f3bb41506db1066dcd1c51bb0f4b02
Tokenizer	meta-llama/Llama-3.1-8B
Data	openai/gsm8k main/train, one ordered row per update, no shuffle
Sequence	[BOS] + body[:510] + [EOS], max 512 tokens
Targets	input_ids = ids[:-1], targets = ids[1:]
Seed	1234
Precision	bfloat16
Optimizer	AdamW, lr 2e-4, weight decay 0.01, no scheduler
Clipping	global norm 1.0
Observation cadence	every 100th epoch-local update
Cameras	1–6 at every observation point

Each training update is one forward pass, one mean next-token cross-entropy, one backward pass, one global gradient clip, and one optimizer step. Observation points are equally spaced and fixed in advance; they were not selected from loss or gradient anomalies.

Camera families

Camera	Family	NPZ files per observation	Recorded surface
1	time_retention	2	pre/post recurrence-control surfaces
2	pressure_map	1	activation gradients at the in_proj control regions
3	optimizer_delivery	1	parameter gradients, clipping, AdamW state, and realized delta
4	memory_path	2	mixer boundaries, decode-cache summaries, and source-derived quantities
5	embedding_learning	1	exact bf16 state of the tied embedding/LM-head table
6	layer_transaction	2	layer boundary values and gradients, pre/post movement

Layout

observed-runs/<run timestamp>/
├── run_summary.json                  # complete runs only
├── e1_s000100/
│   ├── camera-1/
│   ├── camera-2/
│   ├── camera-3/
│   ├── camera-4/
│   ├── camera-5/
│   └── camera-6/
└── ...

Each camera directory contains its phase NPZ file or files plus:

manifest.json       pins, geometry, tensor records, and semantics
storage_report.json byte accounting by quantity
SHA256SUMS          integrity hashes for the camera bundle

Verify any camera bundle from inside its directory:

shasum -a 256 -c SHA256SUMS

Reading constraints carried in the data

• Below-resolution bf16 movement is recorded as exact zero and reported as a fraction; it is not imputed to a small nonzero value.
• Undefined derived values are masked rather than zero-filled. For example, a cosine against a zero-norm vector is NaN with a paired validity mask.
• backbone.embedding.weight and lm_head.weight are one tied parameter with two computational roles, not two independent tensors.
• memory_path keeps live boundary values, official decode-cache summaries, and source-derived quantities as separate evidence classes. Its cache arrays are magnitude and direction summaries only. They do not establish what a token retains or for how long, and state_claims_enabled remains false.
• Camera timing includes observation overhead and is not a corrected estimate of uninstrumented training speed.

Scope

Recorded measurements only. Array names are not interpretations, and the retained partial replication must not be treated as a completed second run.