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

Model ID	`reaperdoesntknow/MoA-100M`
Architecture	`moa_metric` (custom)
Tokenizer	GPT‑2 (`gpt2`) – `pad_token` set to `eos_token`
Context length	2048 tokens
Training data	2 × ≈ 256 k tokens from the datasets listed above
Training compute	CPU‑only (Intel), FP32
Training hyper‑parameters	LR = 5e‑4 (AdamW), batch = 4, seq ≤ 512, 500 k total tokens
Final loss	≈ 0.30 (train)
License	Apache‑2.0
Safety	No alignment or safety fine‑tuning – outputs may be biased or inaccurate.
Intended use	Research on geometry‑aware attention, structured sparsity, and mixture‑of‑attention models.
Limitations	• No KV‑cache → slower generation. • Small token budget → not a general‑purpose LM. • No safety/alignment training.
Out‑of‑scope	High‑stakes applications (medical, legal, etc.) without further evaluation.

Overview

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

Head type	Description
LocalConvHead	Depthwise‑separable 1‑D convolution → captures short‑range context.
Metric Multi‑Head Attention (MetricMHAttention)	Soft‑min over L2 / cosine / diagonal‑Mahalanobis distances: (\displaystyle \text{attn}_{h}(i,j) \propto \exp!\big(-\alpha_h\|q_i-k_j\|^2\big))
Metric MQA	Multi‑Query attention (shared K/V) in the same metric space – cheaper than full MHA.
ChannelMixHead	Per‑token MLP that mixes channel dimensions (no positional mixing).

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)

json

{
  "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)

python

>>> 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)

python

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

Feature	What it does
Ball pruning	Masks keys that lie outside a learned radius → reduces the quadratic attention cost.
Metric MQA	Shares K/V across heads → fewer projection matrices, lower FLOPs.
HyperFFN branch router	Token‑wise top‑k routing means only the most useful branch is evaluated per token.
CPU tips	Set `OMP_NUM_THREADS` / `MKL_NUM_THREADS` to the number of physical cores; use `torch.set_num_threads()` if needed.

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

bibtex

@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

Version	Date	Notes
v0.2	2025‑09‑20	500 k‑token CPU run, GPT‑2 tokenizer, LR = 5e‑4, final loss ≈ 0.30.
v0.1	2025‑09‑20	Initial public release: metric heads, MQA, ball pruning, HyperFFN, router & gates; HF‑compatible; no KV cache.

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

Model	Downloads	Format
MoA-150M	4	HF
MoA-155M	2	HF
MoA-400M	3	HF

Top Models from Our Lab

Model	Downloads
Qwen3-1.7B-Thinking-Distil	501
LFM2.5-1.2B-Distilled-SFT	342
Qwen3-1.7B-Coder-Distilled-SFT	302
Qwen3-0.6B-Distilled-30B-A3B-Thinking-SFT-GGUF	203
Qwen3-1.7B-Coder-Distilled-SFT-GGUF	194

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