MoAMetricLM‑100M — Mixture of Attentions (MoA)

A geometry‑aware Transformer that mixes several attention mechanisms and routes them with a metric‑based router.

• Parameters: ~185 M (≈ 100 M effective due to the mixture)
• Task: Causal language modeling (decoder‑only)
• Library: 🤗 Transformers
• KV cache: Not yet implemented (generation recomputes the full context at every step)

Model card

Overview

MoA replaces the classic dot‑product attention with metric‑based attention and blends four distinct heads per Transformer block:

A token‑wise router decides, for each token, which head(s) to use and applies feature‑gates (FiLM‑style) and router‑bias gates for up/down‑scaling.

The FFN is a HyperFFN – three parallel branches (SwiGLU MLP, separable‑conv, low‑rank) combined by a branch router. LayerScale and optional DropPath keep training stable.

Regularisation (optional)

• Triangle‑inequality (TI) penalty on sampled triples to encourage true‑metric behaviour.
• Ball pruning – each head learns an origin (o_h) and radius (r_h); keys outside the ball are masked, giving structured sparsity.

Architecture diagram (high‑level)

Input → Embedding → (PreNorm) → Block₁ → … → Blockₙ → LM‑Head → Output
                     │
                     ├─ LocalConvHead
                     ├─ MetricMHAttention
                     ├─ MetricMQA
                     └─ ChannelMixHead
                     (router decides per‑token)

Each Block also contains:
  → HyperFFN (SwiGLU | Conv | Low‑rank)  ← branch router
  → LayerScale + DropPath

Configuration (example)

{
  "model_type": "moa_metric",
  "vocab_size": 50257,
  "dim": 768,
  "num_layers": 12,
  "attn_heads": 8,
  "mqa_q_heads": 8,
  "mixer_hidden": 3072,
  "ffn_hidden": 3072,
  "metric": "l2",                     // "l2" | "cosine" | "maha_diag"
  "alpha_init": 1.0,
  "learn_alpha": true,
  "use_balls": true,
  "radius_init": 3.0,
  "learn_radius": true,
  "origin_init_scale": 0.0,
  "maha_init": 1.0,
  "ti_reg_weight": 0.0,
  "ti_reg_samples": 0,
  "router_hidden": 128,
  "router_dropout": 0.1,
  "router_temperature": 1.0,
  "attn_drop": 0.1,
  "proj_drop": 0.1,
  "drop_path": 0.0,
  "max_position_embeddings": 2048,
  "pad_token_id": 50256,
  "bos_token_id": 50256,
  "eos_token_id": 50256
}

Tip: If you use the GPT‑2 tokenizer, set pad_token = eos_token and make sure vocab_size matches the tokenizer (50257).

Quick‑start (inference)

>>> from transformers import AutoTokenizer, AutoModelForCausalLM

>>> model_id = "reaperdoesntknow/MoA-100M"
>>> tokenizer = AutoTokenizer.from_pretrained(model_id)
>>> tokenizer.pad_token = tokenizer.eos_token   # needed for the GPT‑2 tokenizer

>>> model = AutoModelForCausalLM.from_pretrained(model_id)

>>> prompt = "Explain metric‑based attention in simple terms:"
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> output_ids = model.generate(
...     **inputs,
...     max_new_tokens=128,
...     do_sample=False,          # deterministic; set temperature>0 for sampling
...     pad_token_id=tokenizer.pad_token_id,
... )
>>> print(tokenizer.decode(output_ids[0], skip_special_tokens=True))

Note: Because KV‑cache is not implemented, generation time grows linearly with the total context length.

Training (custom loop sketch)

from transformers import AutoTokenizer, AutoModelForCausalLM, DataCollatorForLanguageModeling
from torch.utils.data import DataLoader
import torch, torch.nn.functional as F

tokenizer = AutoTokenizer.from_pretrained("gpt2")
tokenizer.pad_token = tokenizer.eos_token

def collate_fn(examples):
    batch = tokenizer(
        [ex["text"] for ex in examples],
        padding="max_length",
        truncation=True,
        max_length=512,
        return_tensors="pt",
    )
    labels = batch["input_ids"].clone()
    labels[batch["attention_mask"] == 0] = -100
    batch["labels"] = labels
    return batch

# dataset = load_dataset(..., split="train")  # must contain a 'text' field
# loader = DataLoader(dataset, batch_size=4, shuffle=True, collate_fn=collate_fn)

model = AutoModelForCausalLM.from_pretrained("reaperdoesntknow/MoA-100M")
optimizer = torch.optim.AdamW(
    model.parameters(),
    lr=5e-4,
    betas=(0.9, 0.95),
    weight_decay=0.01,
)

for batch in loader:
    out = model(**batch)
    out.loss.backward()
    torch.nn.utils.clip_grad_norm_(model.parameters(), 1.2)
    optimizer.step()
    optimizer.zero_grad()

Evaluation checklist

• Perplexity on a held‑out split of the two training datasets.
• Ablation studies (keep total token budget constant):
  • L2 vs. cosine vs. diagonal‑Mahalanobis distance.
  • With / without ball pruning.
  • With / without HyperFFN branch router.
  • With / without TI regulariser.
• Speed / memory comparison against a vanilla GPT‑2‑size model (same dim/layers).

Efficiency notes

Future roadmap: metric‑aware KV‑cache, kernelised distance approximations (e.g., Random Fourier Features), quantisation & mixed‑precision inference.

Safety, Bias & Risks

• The model has not been fine‑tuned for safety or alignment.
• Outputs may contain biases, profanity, or factual errors.
• Do not deploy in high‑stakes contexts without additional evaluation, moderation, and possibly further fine‑tuning.

Discrepancy Calculus Foundation

This model is part of the Convergent Intelligence LLC: Research Division portfolio. All models in this portfolio are developed under the Discrepancy Calculus (DISC) framework — a measure-theoretic approach to understanding and controlling the gap between what a model should produce and what it actually produces.

DISC treats training singularities (loss plateaus, mode collapse, catastrophic forgetting) not as failures to be smoothed over, but as structural signals that reveal the geometry of the learning problem. Key concepts:

• Discrepancy Operator (D): Measures the gap between expected and observed behavior at each training step
• Jump Sets: Boundaries where model behavior changes discontinuously — these are features, not bugs
• Ghost Imprinting: Teacher knowledge that transfers to student models through weight-space topology rather than explicit distillation signal

For the full mathematical treatment, see Discrepancy Calculus: Foundations and Core Theory (DOI: 10.57967/hf/8194).

Citation chain: Structure Over Scale (DOI: 10.57967/hf/8165) → Three Teachers to Dual Cognition (DOI: 10.57967/hf/8184) → Discrepancy Calculus (DOI: 10.57967/hf/8194)

License

Apache‑2.0 – see the LICENSE file in the repository.

Citation

@misc{moametriclm185m,
  title   = {reaperdoesntknow/MoA-100M: A Geometry-Aware Mixture-of-Attentions Language Model},
  author  = {Convergent Intelligencehawn and collaborators},
  year    = {2025},
  url     = {https://huggingface.co/reaperdoesntknow/MoA-100M}
}

Changelog

Maintainers

• Author: reaper (Convergent Intelligence LLC)
• Contact: Email (convergentintelligencenyc@gmail.com)*

Special Remarks

• This models still in an extremely experimental state. As are most of them, but im working on stabilizing this one for general inference.
• I design create and train all of my models using my mathematical research and pure disgust for the dot product!
• For those of you who actually read this and use my models, you make my day everytime I see another download, so thank you for being awesome!

Convergent Intelligence Portfolio

Part of the Mixture of Attention Series by Convergent Intelligence LLC: Research Division

Related Models

Top Models from Our Lab

Total Portfolio: 41 models | 2,781 total downloads

Last updated: 2026-03-28 12:56 UTC

From the Convergent Intelligence Portfolio

DistilQwen Collection — Our only BF16 series. Proof-weighted distillation from Qwen3-30B-A3B → 1.7B and 0.6B on H100. Three teacher variants (Instruct, Thinking, Coder), nine models, 2,788 combined downloads. The rest of the portfolio proves structure beats scale on CPU. This collection shows what happens when you give the methodology real hardware.

Top model: Qwen3-1.7B-Coder-Distilled-SFT — 508 downloads

Full methodology: Structure Over Scale (DOI: 10.57967/hf/8165)

Convergent Intelligence LLC: Research Division