beerus-b200-checkpoints

Beerus: Swarm-Cortex Head + Frontier30 Graph

Beerus is an experimental byte-level recurrent model coupled to a persistent coherence graph. Beerus is the receiving model and answer owner. It is not BulmaX: any BulmaX-derived material named in an artifact is source provenance, not an identity, persona, or model transfer.

This repository is an artifact store rather than a drop-in transformers.AutoModel package. The current durable Beerus system has three separately versioned layers:

1. the native Beerus neural head;
2. the immutable Frontier30 graph state;
3. versioned graph overlays and, after promotion, a small graph-reader adapter.

Current shape

Native Beerus head

The current durable anchor is beerus_qwen_distill_step000684020.safetensors. Its serialized model topology, rather than inherited generic trainer fields, is the authority for the shape below.

Major parameter-bearing components include a 4,128,768-parameter convolution front end, a 3,543,552-parameter recurrent memory, 50,430,209 parameters in the cell mutators, and separate 198,912-parameter byte embedding and readout matrices. These values describe the head; they do not include graph nodes or edges.

The matching checkpoint manifest contains the named tensors, topology, optimizer state, data cursor, and RNG state needed by the custom Beerus runtime.

Frozen Frontier30 graph

The immutable base is beerus_frontier30_extracted_state.state.safetensors.

The safetensors file includes the graph itself, not merely a summary:

It also carries scalar integrity and topology tensors. The public manifest binds the file size, checksum, node count, edge count, orders, and immutability rules. Geometry, edge targets, and base pheromones are frozen; new capability is added through reversible sidecars rather than reshaping this base.

Capability overlay

The first public capability sidecar is beerus_frontier30_capability_transfer_overlay_v1.safetensors.

The sidecar contains no raw source payload and no identity conditioning. Its source lineage is recorded as a BulmaX bridge corpus, while receiving_model, answer_owner, and runtime identity remain Beerus.

What has and has not been proved

The overlay passed held-out graph-only screening. At the selected strength:

The shuffled-overlay test NLL was 2.350196, supporting that the gain depends on the learned graph associations rather than added score mass alone. Text, audio, and vision source strata passed the eligible graph-only gates. These are source-stratified byte-graph results, not proof that Beerus can directly see, hear, or operate a robot. Mixed and robotics strata lacked enough held-out rows for a claim.

The durable step-684020 head has not yet been promoted as a trained Frontier30 consumer. A 52,801-parameter order-aware reader is staged for a controlled campaign with the head and graph frozen. Until real-graph, shuffled-graph, and graph-off controls pass, the graph results must not be reported as end-to-end Beerus reasoning or runtime memory capability.

Architecture boundary

raw bytes
  -> embedding + causal convolution + recurrent memory
  -> Beerus locally regulated swarm cortex
  -> byte readout

frozen Frontier30 graph
  -> order-aware retrieval
  -> versioned overlay
  -> small gated reader (pending trained promotion)
  -> Beerus swarm input

The head checkpoint, graph base, overlay, and future adapter are independent rollback units. The graph is never counted as neural parameters, and an overlay does not overwrite the frozen base.

Intended use

Beerus is research software for byte-level recurrent modeling, swarm routing, persistent graph memory, controlled graph-conditioned generation, and CPU/CUDA/Intel-XPU runtime experiments. Loading requires the custom Beerus source and exact artifact manifests.

Limitations

• The current durable head is still a research checkpoint, not a polished chat model.
• Graph-only gains do not establish end-to-end head use of the graph.
• A large edge count is memory capacity, not a parameter-count or intelligence equivalence claim.
• Safe online writes require filtering, provenance, rollback, and contamination controls.
• Historical checkpoints elsewhere in this repository remain lineage records; they should not silently replace the current anchor.

Reproducibility

• Head manifest: step 684020 manifest
• Frozen graph manifest: graph_base_v1.public_manifest.json
• Overlay package: PUBLIC_PACKAGE.json
• Combined held-out evaluation: heldout_graph_eval_v1.json
• Source-stratified evaluation: heldout_graph_eval_by_modality_v1.json
• Private source repository: https://github.com/dakuwonmoody-lab/beerus

Beerus is under active development. The checksummed artifacts and manifests, not a moving training process, define the durable shape documented here.