Upload README.md with huggingface_hub

README.md

@@ -1,78 +1,525 @@
 ---
-library_name: transformers
+language: en
+license: apache-2.0
 tags:
-- generated_from_trainer
-model-index:
-- name: NeoLLM
-  results: []
+- causal-lm
+- research
+- bf16
+- attention
+- normalization
+- neollm
+datasets:
+- HuggingFaceFW/fineweb-edu
 ---
 
-<!-- This model card has been generated automatically according to the information the Trainer had access to. You
-should probably proofread and complete it, then remove this comment. -->
-
 # NeoLLM
 
-This model is a fine-tuned version of [](https://huggingface.co/) on an unknown dataset.
-It achieves the following results on the evaluation set:
-- Loss: 4.8999
-- Ce: 3.8699
-- Valid Targets: 2083162.0
-- Ppl: 47.9379
-- Mile Loss: 4.4612
-- Mile Ppl: 86.5885
-- Top1 Accuracy: 0.3130
-- Top3 Accuracy: 0.4684
-- Topk Valid Targets: 2083162.0
-- Ntp Loss: 4.4625
-- Ntp Ce Unweighted: 3.8704
-- Mile Reweighting Delta: 0.5912
-- Mu Loss: 0.0009
-- Tweo Loss: 0.0456
-- Nitp Loss: 0.4369
-- Total Model Loss: 4.8999
-- Tweo Weighted Loss: 0.0005
-- Nitp Weighted Loss: 0.4369
-- Auxiliary Loss: 0.4374
-- Regularization Loss: 0.0014
-
-## Model description
-
-More information needed
-
-## Intended uses & limitations
-
-More information needed
-
-## Training and evaluation data
-
-More information needed
-
-## Training procedure
-
-### Training hyperparameters
-
-The following hyperparameters were used during training:
-- learning_rate: 0.001
-- train_batch_size: 64
-- eval_batch_size: 64
-- seed: 42
-- optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
-- lr_scheduler_type: linear
-- num_epochs: 1
-
-### Training results
-
-| Training Loss | Epoch | Step  | Validation Loss | Ce     | Valid Targets | Ppl      | Mile Loss | Mile Ppl | Top1 Accuracy | Top3 Accuracy | Topk Valid Targets | Ntp Loss | Ntp Ce Unweighted | Mile Reweighting Delta | Mu Loss | Tweo Loss | Nitp Loss | Total Model Loss | Tweo Weighted Loss | Nitp Weighted Loss | Auxiliary Loss | Regularization Loss |
-|:-------------:|:-----:|:-----:|:---------------:|:------:|:-------------:|:--------:|:---------:|:--------:|:-------------:|:-------------:|:------------------:|:--------:|:-----------------:|:----------------------:|:-------:|:---------:|:---------:|:----------------:|:------------------:|:------------------:|:--------------:|:-------------------:|
-| 7.2644        | 0.32  | 5000  | 7.4239          | 6.6627 | 2083162.0     | 782.6802 | 6.7177    | 826.9295 | 0.0792        | 0.1572        | 2083162.0          | 6.7180   | 6.6627            | 0.0550                 | 0.0003  | 2.1997    | 0.6840    | 7.4239           | 0.0220             | 0.6840             | 0.7060         | 0.0223              |
-| 7.0508        | 0.64  | 10000 | 7.2002          | 6.4465 | 2083162.0     | 630.5065 | 6.5119    | 673.0709 | 0.0934        | 0.1711        | 2083162.0          | 6.5126   | 6.4465            | 0.0653                 | 0.0008  | 0.3118    | 0.6844    | 7.2002           | 0.0031             | 0.6844             | 0.6876         | 0.0039              |
-| 5.3104        | 0.96  | 15000 | 4.9266          | 3.9072 | 2083162.0     | 49.7573  | 4.4882    | 88.9567  | 0.3094        | 0.4638        | 2083162.0          | 4.4895   | 3.9076            | 0.5809                 | 0.0009  | 0.0500    | 0.4366    | 4.9266           | 0.0005             | 0.4366             | 0.4371         | 0.0014              |
-| 5.2797        | 1.0   | 15625 | 4.8999          | 3.8699 | 2083162.0     | 47.9379  | 4.4612    | 86.5885  | 0.3130        | 0.4684        | 2083162.0          | 4.4625   | 3.8704            | 0.5912                 | 0.0009  | 0.0456    | 0.4369    | 4.8999           | 0.0005             | 0.4369             | 0.4374         | 0.0014              |
-
-
-### Framework versions
-
-- Transformers 5.20.2
-- Pytorch 2.14.1+cu130
-- Datasets 4.8.5
-- Tokenizers 0.23.3
+NeoLLM is a **84.57 M parameter** decoder-only language model trained from scratch on
+[FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) with
+**BF16** compute, completing training in approximately
+**~0h 46m** on **NVIDIA RTX PRO 6000 Blackwell Server Edition**.
+It integrates a collection of recently published attention and normalization techniques
+into a single architecture, with the goal of studying how they interact during
+pretraining. The model is actively being developed and the current checkpoint represents
+an intermediate training state.
+
+> **Author / contact:** [@Kyokopom](https://x.com/Kyokopom) on X
+> **Repository:** [KitsuVp/NeoLLM](https://huggingface.co/KitsuVp/NeoLLM)
+
+---
+
+## Architecture
+
+NeoLLM is a decoder-only transformer with the following configuration:
+
+| Parameter | Value |
+|---|---|
+| Hidden size | 512 |
+| Layers | 12 |
+| Attention heads | 8 |
+| KV heads (GQA) | 4 |
+| Head dim | 64 |
+| Intermediate size | 1536 |
+| Vocabulary | LiquidAI/LFM2.5-1.2B-Thinking tokenizer (64,402 tokens) |
+| Context length | 512 tokens |
+| Leviathan | enabled=True; d_seed=128, modes=8, knots=16 |
+| JTok / JTok-M | enabled=False / False; Torch continuous surfaces, experts=5, top_k=2 |
+| DAGFormer | enabled=False; {'predictor_dim': 256, 'predictor_layers': 2, 'predictor_heads': 4, 'local_dim': 128, 'input_source': 'tokens'} |
+
+### Parameter breakdown
+
+| Parameter bucket | Count |
+|---|---|
+| **Total parameters** | 84.57M (84,567,992) |
+| **Embedding parameters** (tied) | 32.97M (32,973,824) |
+| **Non-embedding parameters** | 51.59M (51,594,168) |
+| **Effective trainable parameters** | 84.57M (84,567,992) |
+
+> Weight tying is **enabled**: the input embedding matrix and the language-model head
+> share the same parameters, so the effective trainable budget is
+> `total − embed = 51.59M`.
+
+### Integrated techniques
+
+NeoLLM combines architecture modules, optional auxiliary objectives, and
+training-time optimizer/stability components from the following papers.
+
+**Embedding and token representation**
+
+- **Learnable Multipliers** ([arXiv:2601.04890](https://arxiv.org/abs/2601.04890)) — Adds
+  per-row and per-column learnable scalar parameters to selected matrix layers and, when
+  enabled, embeddings.
+- **Leviathan** ([arXiv:2601.22040](https://arxiv.org/abs/2601.22040)) — Optional
+  continuous token embedding generator that can replace the discrete input lookup table.
+- **Leviathan-JTok / JTok-M** ([arXiv:2602.00800](https://arxiv.org/abs/2602.00800)) —
+  Optional continuous adaptation of JTok over the shared Leviathan coordinate. JTok
+  applies one normalized surface as a multiplicative MLP gate; JTok-M routes a
+  Sigmoid+TopK mixture of continuous surfaces and adds the paper's depth-scaled
+  residual. Runtime ``jtok_kernel_backend`` selects the PyTorch reference or the
+  CCE Triton adapter; the JTok surfaces remain outside MXFP8 conversion.
+- **KHRONOS** ([arXiv:2505.13315](https://arxiv.org/abs/2505.13315)) — Kernel/basis
+  reference used by the Leviathan continuous token generator implementation.
+- **Spelling Bee Embeddings** ([arXiv:2601.18030](https://arxiv.org/abs/2601.18030)) —
+  Augments token embeddings with character-level spelling information.
+- **Token Embedding Manifold analysis** ([arXiv:2504.01002](https://arxiv.org/abs/2504.01002)) —
+  Reference motivation for treating token embeddings as structured objects rather than
+  unconstrained lookup rows.
+
+**Attention, positions, and output projection**
+
+- **FAN** ([arXiv:2502.21309](https://arxiv.org/abs/2502.21309)) — Fourier Analysis Networks.
+  A portion of the projection channels are dedicated to periodic cosine/sine features.
+- **MEA** ([arXiv:2601.19611](https://arxiv.org/abs/2601.19611)) — Explicit Multi-head
+  Attention. Adds small learnable interaction matrices between attention heads for K and V.
+- **LUCID** ([arXiv:2602.10410](https://arxiv.org/abs/2602.10410)) — Applies a learned
+  lower-triangular preconditioner to V before attention, decorrelating value representations
+  across positions.
+- **Affine-Scaled Attention** ([arXiv:2602.23057](https://arxiv.org/abs/2602.23057)) — Adds
+  two learnable per-head scalars (α and β) to the softmax weights:
+  `[α·softmax(QKᵀ) + β]·V`.
+- **XSA** ([arXiv:2603.09078](https://arxiv.org/abs/2603.09078)) — Exclusive Self Attention.
+  After computing attention, removes the component of the output aligned with the token's
+  own value vector.
+- **Directional Routing** ([arXiv:2603.14923](https://arxiv.org/abs/2603.14923)) — Each head
+  learns K=4 directions in the output space; a learned router suppresses the attention output
+  along each direction per input.
+- **Gated Attention** ([arXiv:2505.06708](https://arxiv.org/abs/2505.06708)) — A sigmoid gate
+  is applied to the attention output before the output projection, introducing non-linearity
+  and preventing attention sinks.
+- **Momentum Attention** ([arXiv:2411.03884](https://arxiv.org/abs/2411.03884)) — Modifies Q
+  and K by subtracting a fraction of the previous position's Q and K values (causal
+  first-difference).
+- **Interleaved Head Attention / IHA** ([arXiv:2602.21371](https://arxiv.org/abs/2602.21371)) —
+  Builds pseudo-heads from learned cross-head mixtures to create multiple attention patterns
+  per original head.
+- **REPO** ([arXiv:2512.14391](https://arxiv.org/abs/2512.14391)) — Context re-positioning
+  module that learns contextual position coordinates above a configurable start layer.
+- **GRAPE** ([arXiv:2512.07805](https://arxiv.org/abs/2512.07805)) — Group representational
+  position encoding used by the REPO-GRAPE positional path.
+- **GOAT priors** ([arXiv:2601.15380](https://arxiv.org/abs/2601.15380)) — Optional
+  factorized attention log-prior channels inspired by trainable attention priors.
+- **Hadamard output projection** ([arXiv:2603.08343](https://arxiv.org/abs/2603.08343)) —
+  Replaces dense attention output projection with a structured Hadamard transform plus
+  lightweight scaling.
+
+**Normalization, residual flow, and MLP**
+
+- **SeeDNorm** ([arXiv:2510.22777](https://arxiv.org/abs/2510.22777)) — Applied to Q and K
+  projections. Dynamically rescales normalization from the input's own statistics.
+- **LayerNorm Scaling / LNS** ([arXiv:2502.05795](https://arxiv.org/abs/2502.05795)) — Each
+  layer's output is scaled by 1/√ℓ where ℓ is the layer index.
+- **GPAS** ([arXiv:2506.22049](https://arxiv.org/abs/2506.22049)) — Gradient-Preserving
+  Activation Scaling for residual junctions.
+- **PolyNorm** ([arXiv:2602.04902](https://arxiv.org/abs/2602.04902)) — Replaces the standard
+  MLP activation with normalized linear, quadratic, and cubic branches.
+- **SimpleGPT** ([arXiv:2602.01212](https://arxiv.org/abs/2602.01212)) — Second-order
+  geometry-inspired normalization strategy applied inside MLP projections.
+- **StackMemory / STACKTRANS** ([NeurIPS 2025](https://openreview.net/forum?id=2bbDg587uh)) —
+  Optional differentiable hidden-state stack between decoder layers.
+- **Attention Residuals / AttnRes** ([arXiv:2603.15031](https://arxiv.org/abs/2603.15031)) —
+  Optional learned depth-wise aggregation over previous layer outputs or block summaries.
+- **LAUREL** ([arXiv:2411.07501](https://arxiv.org/abs/2411.07501)) — Optional learned
+  augmented residual layer with residual-weight and low-rank variants.
+
+**Training objectives and training-time regularizers**
+
+- **Cut Cross Entropy** ([Apple repository](https://github.com/apple/ml-cross-entropy)) —
+  Memory-efficient next-token loss that avoids materializing the full token-by-vocabulary
+  logits tensor. NeoLLM remains compatible with the upstream package when the extensions
+  below are disabled.
+- **MiLe Loss** ([arXiv:2310.19531](https://arxiv.org/abs/2310.19531)) — Optional detached,
+  mean-normalized predictive-entropy weighting of token losses, implemented inside the
+  extended CCE path.
+- **Output Embedding Centering / mu-loss**
+  ([arXiv:2601.02031](https://arxiv.org/abs/2601.02031)) — Optional
+  `lambda * ||mean(output_embeddings)||^2` regularizer for output-logit stability.
+- **MEAP** ([arXiv:2502.07490](https://arxiv.org/abs/2502.07490)) — Optional training-only
+  input corruption that masks a fixed fraction of eligible tokens while preserving clean
+  next-token labels, causal attention, and the inference path.
+- **TWEO** ([arXiv:2511.23225](https://arxiv.org/abs/2511.23225)) — Optional
+  Transformers Without Extreme Outliers activation regularizer for FP8/low-bit-friendly
+  training.
+- **NITP** ([arXiv:2605.24956](https://arxiv.org/abs/2605.24956)) — Optional Next Implicit
+  Token Prediction auxiliary objective using shallow-layer implicit token targets and a
+  cosine loss.
+- **NextLat** ([arXiv:2511.05963](https://arxiv.org/abs/2511.05963)) — Optional next-latent
+  prediction objective using latent dynamics, Smooth L1 supervision, and frozen-head KL.
+
+### Optional extended-CCE configuration
+
+| Feature | Enabled | Value |
+|---|---:|---:|
+| MiLe Loss | True | gamma=1.0 |
+| mu-loss | True | lambda=0.0001 |
+| MEAP | True | ratio=0.15 |
+
+MiLe, mu-loss, and MEAP require the extended
+[`Kitsunp/ml-cross-entropy`](https://github.com/Kitsunp/ml-cross-entropy) package only when
+their corresponding flags are enabled. With all flags disabled, NeoLLM calls upstream CCE
+without extension-specific arguments. When any extension is active, CCE reports three compact
+scalars: unweighted NTP cross entropy, the MiLe reweighting delta, and the mu-loss penalty.
+Their sum reconstructs `ntp_loss` exactly. MEAP reports its eligible and selected counts from
+the masking kernel; the trainer logs the selected count and exact fraction. No diagnostic path
+materializes full-vocabulary logits or a token mask outside the kernels.
+
+**Optimizer and training stability**
+
+- **Adam-family optimizer mode** — Active variant: **AdEMAMix**.
+  **AdEMAMix** ([arXiv:2409.03137](https://arxiv.org/abs/2409.03137)) contributes the optional
+  beta3 slow-memory branch when `ADEMAMIX=True`. **DeltaMomentum**
+  ([arXiv:2608.19491](https://arxiv.org/abs/2608.19491)) optionally replaces only the fast beta1
+  first-moment update on hidden Linear weights when `DELTA=True`; beta2 remains Adam's
+  ordinary squared-gradient EMA in every mode.
+- **BF16 stochastic rounding** ([arXiv:2502.20566](https://arxiv.org/abs/2502.20566)) —
+  Optional master-free BF16 model update with stochastic rather than nearest rounding; the same
+  write-back path is shared by AdamW, DeltaAdamW, AdEMAMix, and DeltaAdEMAMix.
+- **Cautious Weight Decay** ([arXiv:2510.12402](https://arxiv.org/abs/2510.12402)) —
+  Sign-selective weight decay variant used by the custom optimizer logic.
+- **Correction of Decoupled Weight Decay** ([arXiv:2512.08217](https://arxiv.org/abs/2512.08217)) —
+  Adapts decoupled weight decay during learning-rate decay.
+- **AdamHD** ([arXiv:2511.14721](https://arxiv.org/abs/2511.14721)) —
+  Decoupled Huber decay regularization reference used by the optimizer.
+- **Gradient conditioning** — CAST active; history-adaptive spectral/norm policies by optimizer group; 15625 committed updates.
+  References: [CAST](https://arxiv.org/abs/2610.19761),
+  [GradientStabilizer](https://arxiv.org/abs/2502.17055).
+---
+
+## Training
+
+| Setting | Value |
+|---|---|
+| Dataset | FineWeb-Edu (sample-10BT) |
+| Tokens seen | ~0.51B (15,625 steps × batch 64 × length 512) |
+| Precision | BF16 model state (bfloat16: 84,565,816 elements, float32: 2,176 elements) + AdEMAMix FP32 master active; stochastic rounding disabled |
+| Optimizer | AdEMAMix |
+| Gradient transform | CAST active; history-adaptive spectral/norm policies by optimizer group; 15625 committed updates |
+| Learning rate | 1e-03 with 4,000-step linear warmup |
+| Weight decay | coefficients [0.005, 0.01, 0.1]; WD 5/5 groups; CWD 2; Huber 0; Correction 0 |
+| Training time | ~0h 46m |
+| Hardware | NVIDIA RTX PRO 6000 Blackwell Server Edition |
+
+
+### Training curve — total objective
+
+Total includes MiLe when enabled, plus all active auxiliary objectives and regularizers. Without MiLe means model total minus its MiLe delta on the same forward; this is not a separate training run without MiLe.
+
+| Step | Train total | Val total | Train without MiLe | Val without MiLe |
+|---|---|---|---|---|
+| 1 | 12.1388 | — | 12.1388 | — |
+| 250 | 7.98512 | — | 7.97511 | — |
+| 500 | 7.8254 | — | 7.80911 | — |
+| 750 | 7.69152 | — | 7.67302 | — |
+| 1,000 | 7.60931 | — | 7.58873 | — |
+| 1,250 | 7.55672 | — | 7.53484 | — |
+| 1,500 | 7.52617 | — | 7.50214 | — |
+| 1,750 | 7.49842 | — | 7.47334 | — |
+| 2,000 | 7.47787 | — | 7.45123 | — |
+| 2,250 | 7.47165 | — | 7.4431 | — |
+| 2,500 | 7.44367 | — | 7.41406 | — |
+| 2,750 | 7.42427 | — | 7.39387 | — |
+| 3,000 | 7.40542 | — | 7.37407 | — |
+| 3,250 | 7.38995 | — | 7.35702 | — |
+| 3,500 | 7.37661 | — | 7.34258 | — |
+| 3,750 | 7.3544 | — | 7.31952 | — |
+| 4,000 | 7.34423 | — | 7.30894 | — |
+| 4,250 | 7.32617 | — | 7.28946 | — |
+| 4,500 | 7.30903 | — | 7.27187 | — |
+| 4,750 | 7.28931 | — | 7.2511 | — |
+| 5,000 | 7.26439 | 7.42393 | 7.22555 | 7.36892 |
+| 5,250 | 7.24687 | — | 7.20728 | — |
+| 5,500 | 7.23375 | — | 7.19312 | — |
+| 5,750 | 7.21333 | — | 7.17154 | — |
+| 6,000 | 7.1986 | — | 7.15614 | — |
+| 6,250 | 7.18516 | — | 7.14238 | — |
+| 6,500 | 7.1682 | — | 7.12551 | — |
+| 6,750 | 7.15914 | — | 7.11515 | — |
+| 7,000 | 7.1406 | — | 7.09638 | — |
+| 7,250 | 7.13519 | — | 7.09033 | — |
+| 7,500 | 7.12815 | — | 7.08269 | — |
+| 7,750 | 7.1143 | — | 7.06872 | — |
+| 8,000 | 7.10478 | — | 7.05756 | — |
+| 8,250 | 7.09332 | — | 7.04579 | — |
+| 8,500 | 7.09115 | — | 7.04364 | — |
+| 8,750 | 7.09173 | — | 7.04375 | — |
+| 9,000 | 7.08177 | — | 7.03372 | — |
+| 9,250 | 7.06866 | — | 7.01952 | — |
+| 9,500 | 7.06379 | — | 7.01455 | — |
+| 9,750 | 7.06571 | — | 7.01678 | — |
+| 10,000 | 7.05078 | 7.20018 | 7.00036 | 7.13484 |
+| 10,250 | 7.05115 | — | 7.00069 | — |
+| 10,500 | 6.84575 | — | 6.73927 | — |
+| 10,750 | 5.93191 | — | 5.6053 | — |
+| 11,000 | 5.76889 | — | 5.40413 | — |
+| 11,250 | 5.68452 | — | 5.29233 | — |
+| 11,500 | 5.62871 | — | 5.21546 | — |
+| 11,750 | 5.58288 | — | 5.15794 | — |
+| 12,000 | 5.55091 | — | 5.11768 | — |
+| 12,250 | 5.51622 | — | 5.07728 | — |
+| 12,500 | 5.48044 | — | 5.03584 | — |
+| 12,750 | 5.46397 | — | 5.01222 | — |
+| 13,000 | 5.43349 | — | 4.97721 | — |
+| 13,250 | 5.415 | — | 4.95451 | — |
+| 13,500 | 5.39183 | — | 4.92584 | — |
+| 13,750 | 5.38363 | — | 4.91604 | — |
+| 14,000 | 5.35952 | — | 4.88958 | — |
+| 14,250 | 5.34301 | — | 4.87144 | — |
+| 14,500 | 5.32907 | — | 4.8535 | — |
+| 14,750 | 5.31257 | — | 4.83203 | — |
+| 15,000 | 5.31041 | 4.92658 | 4.83019 | 4.34564 |
+| 15,250 | 5.28783 | — | 4.80606 | — |
+| 15,500 | 5.27966 | — | 4.79588 | — |
+| 15,625 | 5.26531 | 4.89987 | 4.78016 | 4.30866 |
+
+### Cross-entropy with and without MiLe
+
+CE without MiLe is unweighted CE; CE with MiLe adds the reweighting delta. Both exclude mu and auxiliary terms. `ntp_loss` includes mu when enabled.
+
+| Step | Train CE without MiLe | Val CE without MiLe | Train CE with MiLe | Val CE with MiLe |
+|---|---|---|---|---|
+| 1 | 11.1797 | — | 11.1797 | — |
+| 250 | 7.70254 | — | 7.71255 | — |
+| 500 | 7.37642 | — | 7.39271 | — |
+| 750 | 7.23382 | — | 7.25233 | — |
+| 1,000 | 7.1488 | — | 7.16938 | — |
+| 1,250 | 7.08809 | — | 7.10998 | — |
+| 1,500 | 7.04248 | — | 7.06651 | — |
+| 1,750 | 6.99841 | — | 7.02348 | — |
+| 2,000 | 6.96604 | — | 6.99268 | — |
+| 2,250 | 6.94518 | — | 6.97373 | — |
+| 2,500 | 6.91234 | — | 6.94194 | — |
+| 2,750 | 6.88875 | — | 6.91915 | — |
+| 3,000 | 6.86703 | — | 6.89838 | — |
+| 3,250 | 6.84514 | — | 6.87807 | — |
+| 3,500 | 6.82863 | — | 6.86265 | — |
+| 3,750 | 6.80252 | — | 6.8374 | — |
+| 4,000 | 6.7882 | — | 6.8235 | — |
+| 4,250 | 6.76929 | — | 6.80601 | — |
+| 4,500 | 6.75193 | — | 6.78909 | — |
+| 4,750 | 6.73117 | — | 6.76937 | — |
+| 5,000 | 6.70779 | 6.66266 | 6.74662 | 6.71767 |
+| 5,250 | 6.69276 | — | 6.73236 | — |
+| 5,500 | 6.68077 | — | 6.72141 | — |
+| 5,750 | 6.66076 | — | 6.70255 | — |
+| 6,000 | 6.64853 | — | 6.69098 | — |
+| 6,250 | 6.6371 | — | 6.67987 | — |
+| 6,500 | 6.62383 | — | 6.66651 | — |
+| 6,750 | 6.61443 | — | 6.65842 | — |
+| 7,000 | 6.5967 | — | 6.64092 | — |
+| 7,250 | 6.59131 | — | 6.63617 | — |
+| 7,500 | 6.58502 | — | 6.63047 | — |
+| 7,750 | 6.57292 | — | 6.6185 | — |
+| 8,000 | 6.56308 | — | 6.61029 | — |
+| 8,250 | 6.55154 | — | 6.59907 | — |
+| 8,500 | 6.54732 | — | 6.59483 | — |
+| 8,750 | 6.54686 | — | 6.59484 | — |
+| 9,000 | 6.53598 | — | 6.58403 | — |
+| 9,250 | 6.52115 | — | 6.57028 | — |
+| 9,500 | 6.51535 | — | 6.56459 | — |
+| 9,750 | 6.51733 | — | 6.56626 | — |
+| 10,000 | 6.50188 | 6.44648 | 6.55229 | 6.51183 |
+| 10,250 | 6.50001 | — | 6.55046 | — |
+| 10,500 | 6.22545 | — | 6.33193 | — |
+| 10,750 | 5.14674 | — | 5.47334 | — |
+| 11,000 | 4.9484 | — | 5.31316 | — |
+| 11,250 | 4.8344 | — | 5.2266 | — |
+| 11,500 | 4.75564 | — | 5.1689 | — |
+| 11,750 | 4.69681 | — | 5.12174 | — |
+| 12,000 | 4.65506 | — | 5.0883 | — |
+| 12,250 | 4.61458 | — | 5.05351 | — |
+| 12,500 | 4.57247 | — | 5.01707 | — |
+| 12,750 | 4.54935 | — | 5.0011 | — |
+| 13,000 | 4.51444 | — | 4.97073 | — |
+| 13,250 | 4.49163 | — | 4.95212 | — |
+| 13,500 | 4.46322 | — | 4.92921 | — |
+| 13,750 | 4.45248 | — | 4.92008 | — |
+| 14,000 | 4.4259 | — | 4.89584 | — |
+| 14,250 | 4.40784 | — | 4.87941 | — |
+| 14,500 | 4.38954 | — | 4.86512 | — |
+| 14,750 | 4.36861 | — | 4.84916 | — |
+| 15,000 | 4.3654 | 3.90764 | 4.84562 | 4.48858 |
+| 15,250 | 4.342 | — | 4.82377 | — |
+| 15,500 | 4.33173 | — | 4.8155 | — |
+| 15,625 | 4.31639 | 3.87037 | 4.80155 | 4.46158 |
+
+### Observed loss components — weighted contributions
+
+Only contributions actually added to the objective are shown; raw values and subcomponents remain in training metrics. `regularization_loss` is the subtotal of mu, weighted TWeo, JTok-M and StackMemory entropy, excluding NITP/NextLat; do not add it again to those components. JTok-M auxiliary loss is training-only.
+
+| Step | Phase | mu | TWeo | NITP | Reg. subtotal |
+|---|---|---|---|---|---|
+| 1 | train | 3.18928e-10 | 8.85304e-07 | 0.959145 | 8.85623e-07 |
+| 250 | train | 5.68541e-06 | 0.000889012 | 0.271676 | 0.000894698 |
+| 500 | train | 1.39317e-05 | 0.00048898 | 0.43219 | 0.000502912 |
+| 750 | train | 2.39475e-05 | 0.000592826 | 0.438579 | 0.000616773 |
+| 1,000 | train | 3.58659e-05 | 0.000741196 | 0.439151 | 0.000777062 |
+| 1,250 | train | 4.88121e-05 | 0.000991387 | 0.445703 | 0.0010402 |
+| 1,500 | train | 6.18302e-05 | 0.00132114 | 0.458276 | 0.00138297 |
+| 1,750 | train | 7.31645e-05 | 0.00163189 | 0.47323 | 0.00170506 |
+| 2,000 | train | 8.32728e-05 | 0.00189371 | 0.483209 | 0.00197699 |
+| 2,250 | train | 9.34909e-05 | 0.00216144 | 0.49566 | 0.00225493 |
+| 2,500 | train | 0.000104415 | 0.00233031 | 0.499293 | 0.00243473 |
+| 2,750 | train | 0.000115916 | 0.00246703 | 0.502539 | 0.00258294 |
+| 3,000 | train | 0.000127831 | 0.00244716 | 0.504458 | 0.002575 |
+| 3,250 | train | 0.000141259 | 0.00240863 | 0.509329 | 0.00254989 |
+| 3,500 | train | 0.0001555 | 0.00235224 | 0.51145 | 0.00250774 |
+| 3,750 | train | 0.000170426 | 0.00234662 | 0.51449 | 0.00251704 |
+| 4,000 | train | 0.000185567 | 0.00222962 | 0.518318 | 0.00241518 |
+| 4,250 | train | 0.000203734 | 0.00213374 | 0.51783 | 0.00233748 |
+| 4,500 | train | 0.00022362 | 0.00194741 | 0.517777 | 0.00217103 |
+| 4,750 | train | 0.000244492 | 0.00176364 | 0.517932 | 0.00200813 |
+| 5,000 | train | 0.000269174 | 0.00161852 | 0.515877 | 0.00188769 |
+| 5,000 | validation | 0.000281882 | 0.0219972 | 0.683979 | 0.0222791 |
+| 5,250 | train | 0.0002948 | 0.00152122 | 0.512702 | 0.00181602 |
+| 5,500 | train | 0.000320074 | 0.00141181 | 0.510612 | 0.00173188 |
+| 5,750 | train | 0.000344429 | 0.00129077 | 0.50915 | 0.0016352 |
+| 6,000 | train | 0.000369502 | 0.00119281 | 0.506049 | 0.00156231 |
+| 6,250 | train | 0.00039563 | 0.00114157 | 0.503747 | 0.0015372 |
+| 6,500 | train | 0.000424725 | 0.00111766 | 0.500143 | 0.00154238 |
+| 6,750 | train | 0.000454083 | 0.00109027 | 0.499179 | 0.00154435 |
+| 7,000 | train | 0.00048038 | 0.00103857 | 0.498165 | 0.00151895 |
+| 7,250 | train | 0.00050527 | 0.000984053 | 0.497533 | 0.00148932 |
+| 7,500 | train | 0.000530754 | 0.000993553 | 0.49615 | 0.00152431 |
+| 7,750 | train | 0.000560454 | 0.000970533 | 0.494275 | 0.00153099 |
+| 8,000 | train | 0.000594531 | 0.000919749 | 0.492966 | 0.00151428 |
+| 8,250 | train | 0.000624764 | 0.000917213 | 0.492705 | 0.00154198 |
+| 8,500 | train | 0.000650141 | 0.00152388 | 0.494143 | 0.00217402 |
+| 8,750 | train | 0.000673515 | 0.000843965 | 0.495374 | 0.00151748 |
+| 9,000 | train | 0.000697932 | 0.000824539 | 0.496217 | 0.00152247 |
+| 9,250 | train | 0.000722145 | 0.000783098 | 0.496869 | 0.00150524 |
+| 9,500 | train | 0.000745877 | 0.000778402 | 0.497681 | 0.00152428 |
+| 9,750 | train | 0.000765973 | 0.000764489 | 0.497919 | 0.00153046 |
+| 10,000 | train | 0.000788166 | 0.000781291 | 0.496911 | 0.00156946 |
+| 10,000 | validation | 0.00080187 | 0.00311829 | 0.684435 | 0.00392016 |
+| 10,250 | train | 0.000816356 | 0.000781903 | 0.499086 | 0.00159826 |
+| 10,500 | train | 0.000843664 | 0.00125388 | 0.511724 | 0.00209754 |
+| 10,750 | train | 0.000872123 | 0.000880954 | 0.456807 | 0.00175308 |
+| 11,000 | train | 0.000892194 | 0.000714178 | 0.454127 | 0.00160637 |
+| 11,250 | train | 0.000900344 | 0.000727329 | 0.456301 | 0.00162767 |
+| 11,500 | train | 0.000902344 | 0.000759315 | 0.45815 | 0.00166166 |
+| 11,750 | train | 0.000908495 | 0.000824227 | 0.459405 | 0.00173272 |
+| 12,000 | train | 0.00091642 | 0.000833896 | 0.460866 | 0.00175032 |
+| 12,250 | train | 0.0009223 | 0.000845974 | 0.460934 | 0.00176827 |
+| 12,500 | train | 0.000924763 | 0.000815765 | 0.46163 | 0.00174053 |
+| 12,750 | train | 0.000923761 | 0.000778682 | 0.461164 | 0.00170244 |
+| 13,000 | train | 0.000920588 | 0.000750611 | 0.461097 | 0.0016712 |
+| 13,250 | train | 0.000917439 | 0.000692365 | 0.461269 | 0.0016098 |
+| 13,500 | train | 0.000913774 | 0.000652368 | 0.461055 | 0.00156614 |
+| 13,750 | train | 0.000910805 | 0.000610711 | 0.462036 | 0.00152152 |
+| 14,000 | train | 0.00090898 | 0.000583299 | 0.462182 | 0.00149228 |
+| 14,250 | train | 0.000907558 | 0.000552937 | 0.462141 | 0.0014605 |
+| 14,500 | train | 0.00090634 | 0.000519922 | 0.462532 | 0.00142626 |
+| 14,750 | train | 0.00090718 | 0.000488421 | 0.462017 | 0.0013956 |
+| 15,000 | train | 0.000909732 | 0.000724183 | 0.463154 | 0.00163391 |
+| 15,000 | validation | 0.000911385 | 0.000499765 | 0.436584 | 0.00141115 |
+| 15,250 | train | 0.000913244 | 0.000435868 | 0.46271 | 0.00134911 |
+| 15,500 | train | 0.000916544 | 0.000419084 | 0.462817 | 0.00133563 |
+| 15,625 | train | 0.000918795 | 0.000406611 | 0.462439 | 0.00132541 |
+| 15,625 | validation | 0.00091954 | 0.000456212 | 0.436919 | 0.00137575 |
+
+Tables use exact logged steps; missing values are — without carrying values forward. Repeated evaluations at one step use the latest logged value. Components average batch scalars weighted by examples; CE/PPL token metrics retain their valid-target weighting. Weight decay/CWD act in the optimizer, outside these forward losses.
+
+---
+
+## Limitations
+
+- **Token budget** — ~1.5 B tokens seen; below estimated optimum. Knowledge-intensive tasks
+  will improve with more training.
+- **Gradient spike at step 40k** — Reorganized the attention pattern in layer 9 that
+  previously captured long-range token correlations. A checkpoint from ~step 38k is expected
+  to have better aggregate benchmark scores.
+- **PolyNorm exclusivity** — The quadratic branch has become partially redundant with the
+  linear branch. Will be corrected in the next training run.
+- **Base model only** — Not instruction-tuned or aligned; purely a next-token-prediction
+  base model.
+
+---
+
+## References
+
+All papers whose techniques are integrated into NeoLLM's architecture,
+training objective, or training stack:
+
+| Area | Technique | Paper title | Reference |
+|---|---|---|---|
+| Embeddings | Learnable Multipliers | Freeing the Scale of Language Model Matrix Layers | [arXiv:2601.04890](https://arxiv.org/abs/2601.04890) |
+| Embeddings | Leviathan | A Separable Architecture for Continuous Token Representation in Language Models | [arXiv:2601.22040](https://arxiv.org/abs/2601.22040) |
+| Embeddings / residual modulation | JTok / JTok-M | JTok: On Token Embedding as another Axis of Scaling Law via Joint Token Self-modulation | [arXiv:2602.00800](https://arxiv.org/abs/2602.00800) |
+| Embeddings | KHRONOS | KHRONOS: a Kernel-Based Neural Architecture for Rapid, Resource-Efficient Scientific Computation | [arXiv:2505.13315](https://arxiv.org/abs/2505.13315) |
+| Embeddings | Spelling Bee | Spelling Bee Embeddings for Language Modeling | [arXiv:2601.18030](https://arxiv.org/abs/2601.18030) |
+| Embeddings | Token embedding analysis | Token Embeddings Violate the Manifold Hypothesis | [arXiv:2504.01002](https://arxiv.org/abs/2504.01002) |
+| Attention / positions | FAN | Fourier Analysis Networks | [arXiv:2502.21309](https://arxiv.org/abs/2502.21309) |
+| Attention / positions | MEA | Explicit Multi-head Attention for Inter-head Interaction in Large Language Models | [arXiv:2601.19611](https://arxiv.org/abs/2601.19611) |
+| Attention / positions | LUCID | Attention with Preconditioned Representations | [arXiv:2602.10410](https://arxiv.org/abs/2602.10410) |
+| Attention / positions | Affine-Scaled Attention | Affine-Scaled Attention: Towards Flexible and Stable Transformer Attention | [arXiv:2602.23057](https://arxiv.org/abs/2602.23057) |
+| Attention / positions | XSA | Exclusive Self Attention | [arXiv:2603.09078](https://arxiv.org/abs/2603.09078) |
+| Attention / positions | Directional Routing | Directional Routing in Transformers | [arXiv:2603.14923](https://arxiv.org/abs/2603.14923) |
+| Attention / positions | Gated Attention | Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free | [arXiv:2505.06708](https://arxiv.org/abs/2505.06708) |
+| Attention / positions | Momentum Attention | Momentum Attention | [arXiv:2411.03884](https://arxiv.org/abs/2411.03884) |
+| Attention / positions | IHA | Interleaved Head Attention | [arXiv:2602.21371](https://arxiv.org/abs/2602.21371) |
+| Attention / positions | REPO | Language Models with Context Re-Positioning | [arXiv:2512.14391](https://arxiv.org/abs/2512.14391) |
+| Attention / positions | GRAPE | Group Representational Position Encoding | [arXiv:2512.07805](https://arxiv.org/abs/2512.07805) |
+| Attention / positions | GOAT priors | You Need Better Attention Priors | [arXiv:2601.15380](https://arxiv.org/abs/2601.15380) |
+| Attention / positions | Hadamard o_proj | Rethinking Attention Output Projection: Structured Hadamard Transforms for Efficient Transformers | [arXiv:2603.08343](https://arxiv.org/abs/2603.08343) |
+| Residual / normalization | SeeDNorm | Self-Rescaled Dynamic Normalization | [arXiv:2510.22777](https://arxiv.org/abs/2510.22777) |
+| Residual / normalization | LNS | The Curse of Depth in LLMs | [arXiv:2502.05795](https://arxiv.org/abs/2502.05795) |
+| Residual / normalization | GPAS | Gradient-Preserving Activation Scaling | [arXiv:2506.22049](https://arxiv.org/abs/2506.22049) |
+| Residual / normalization | PolyNorm | PolyNorm / PolyCom | [arXiv:2602.04902](https://arxiv.org/abs/2602.04902) |
+| Residual / normalization | SimpleGPT | SimpleGPT | [arXiv:2602.01212](https://arxiv.org/abs/2602.01212) |
+| Residual / normalization | StackMemory / STACKTRANS | Recursive Transformer: Boosting Reasoning Ability with State Stack | [NeurIPS 2025](https://openreview.net/forum?id=2bbDg587uh) |
+| Residual / normalization | Attention Residuals | Attention Residuals | [arXiv:2603.15031](https://arxiv.org/abs/2603.15031) |
+| Residual / normalization | LAUREL | LAUREL: Learned Augmented Residual Layer | [arXiv:2411.07501](https://arxiv.org/abs/2411.07501) |
+| Objectives | TWEO | Transformers Without Extreme Outliers Enables FP8 Training And Quantization For Dummies | [arXiv:2511.23225](https://arxiv.org/abs/2511.23225) |
+| Objectives | NITP | Next Implicit Token Prediction for LLM Pre-training | [arXiv:2605.24956](https://arxiv.org/abs/2605.24956) |
+| Objectives | NextLat | Next-Latent Prediction Transformers Learn Compact World Models | [arXiv:2511.05963](https://arxiv.org/abs/2511.05963) |
+| Optimizer / training | AdEMAMix | The AdEMAMix Optimizer: Better, Faster, Older | [arXiv:2409.03137](https://arxiv.org/abs/2409.03137) |
+| Optimizer / training | DeltaMomentum | DeltaMomentum: A Key-Value based Anisotropic Momentum Update via Delta Rule | [arXiv:2608.19491](https://arxiv.org/abs/2608.19491) |
+| Optimizer / training | BF16 + SR | Stochastic Rounding for LLM Training: Theory and Practice | [arXiv:2502.20566](https://arxiv.org/abs/2502.20566) |
+| Optimizer / training | CWD | Cautious Weight Decay | [arXiv:2510.12402](https://arxiv.org/abs/2510.12402) |
+| Optimizer / training | WD correction | Correction of Decoupled Weight Decay | [arXiv:2512.08217](https://arxiv.org/abs/2512.08217) |
+| Optimizer / training | AdamHD | AdamHD: Decoupled Huber Decay Regularization for Language Model Pre-Training | [arXiv:2511.14721](https://arxiv.org/abs/2511.14721) |
+| Optimizer / training | GradientStabilizer | GradientStabilizer | [arXiv:2502.17055](https://arxiv.org/abs/2502.17055) |
+
+---
+
+## Citation
+
+```bibtex
+@misc{neollm2026,
+  title  = {NeoLLM: A Research Language Model Integrating Recent Attention and Normalization Techniques},
+  author = {KitsuVp},
+  year   = {2026},
+  url    = {https://huggingface.co/KitsuVp/NeoLLM}
+}
+```
+
+---
+
+## Author
+
+[@Kyokopom](https://x.com/Kyokopom) on X
+
+---
+
+## License
+
+Apache 2.0