gdn2-mlp-1.3b

GDN2-MLP 1.3B — 152B-token control

This is the matched GDN2-MLP control for Emender E97 1.3B. It is a 1.287B-parameter recurrent base language model trained for 152,280,498,176 tokens. It is not a 152B-parameter model.

The model uses NVIDIA Gated DeltaNet 2 token mixing plus a post-mixer SwiGLU MLP, reshaped by CMA-ES to match the E97 parameter budget. Both controls used the same Pile/p50k_base stream, 2,048-token chunks, BF16 Schedule-Free AdamW, and eight-island DiLoCo harness.

Artifact

• Final step: 2,323,616
• Training-log last-100 loss: 2.4034 nats/token
• Source checkpoint: 7,720,577,595 bytes
• Source checkpoint SHA-256: ffae61cfeeeff820f469f7d66dd31c60e48c455872ff9102e57b15ad75bad59b
• Export: BF16 Schedule-Free train/y weights
• Tokenizer: p50k_base
• Raw pickle checkpoint and optimizer state: not included

At the last common regular log point (150,793,420,800 tokens), the 80-point moving averages were E97 2.437045 and GDN2 2.426705; the difference was -0.010340 nats/token (GDN2 minus E97). These are training-log summaries, not a replacement for fixed held-out evaluation.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("spinozans/gdn2-mlp-1.3b")
model = AutoModelForCausalLM.from_pretrained(
    "spinozans/gdn2-mlp-1.3b", trust_remote_code=True, dtype="auto"
)

The bundled remote code is a standalone portable PyTorch implementation of the published GDN2 recurrence. It does not redistribute NVIDIA source. It favors portability and exact weight loading over fused-kernel throughput.

The original fused GDN2 implementation used for training is available from NVIDIA GatedDeltaNet-2 at commit 95709fc250357c2dd109361c353192f2aa5913f9 under NVIDIA's Source Code License-NC. The Emender training wrapper is recorded at commit a3a862f30f5c3c9584e18490c986bc9a065d6653.

Validation

• all 267 exported tensors match recovered source-checkpoint train/y weights exactly;
• the Transformers loader reports zero missing, unexpected, or mismatched keys;
• portable source-checkpoint and exported CPU logits are bit-identical;
• CUDA fused and portable paths agree on greedy argmax for The theorem states (measured mean/max absolute logit delta 0.01244 / 0.125 from BF16 kernel evaluation order);
• live-style greedy generation produces a finite continuation.

Machine-readable evidence is included in validation.json.

Limitations and license

This is a raw base LM, not an instruction or chat model. It may emit incorrect, biased, or unsafe text. Benchmark coverage is limited. No standalone model license has been selected; license: other is intentional. The separately available NVIDIA fused implementation has its own non-commercial source license, which is not replaced by this model card.