MicroMixer-4-50K-Discord-Dialogues

Micro Language Model
Attention-Free • MLP-Only • Byte-Level • Content-Gated Dilated Convolution

📋 Overview

MicroMixer-4-50K-Discord-Dialogues is a 48,684-parameter pure MLP-Mixer causal language model — no attention, no recurrence, no SSM — pretrained on Discord conversation data and then fine-tuned with FMSP (Fine-tuning with Minimal Parameter changes for Small-parameter LMs) on 9,012 general-knowledge QA pairs.

This is the 50K member of the MicroMixer-4 (V87 Final) family: below the memorization floor — full-988 EM is exactly 0, and fluency is partially degenerate.

The backbone is V87 Final, the project's champion architecture — a CCD-Mixer (Content-gated mixture of shared-weight Dilated convolutions) crowned overall champion of the 1M architecture census (V86), frozen as the final chassis and scaled to six parameter budgets. The 50K preset reproduces the champion recipe verbatim at its budget.

🏗️ Architecture

graph TD
    A[Byte Input] --> B[Embed 256→32 NoPE]
    B --> C[CCD-Mixer Block × 4]
    C --> D[RMSNorm]
    D --> E[LM Head Tied with Embed]
    E --> F[Byte Output]

    subgraph "CCD-Mixer Block"
        X[Input 32] --> U["Linear d→2d → split v, g"]
        U --> RP[Full RoPE on v AND g]
        RP --> M["Shared-weight dilated conv<br/>dilations 1·2·4·8, k=65"]
        M --> G["Per-position 4-way gate<br/>softmax(Linear_dil(x)/τ)"]
        G --> O["W_o(v ⊙ g)  — zero-init"]
        O --> SW[SwiGLU Channel-Mix]
        SW --> RM[ReMixerLayer sidecar]
    end

    style A fill:#007BFF,color:#fff
    style F fill:#00D620,color:#fff
    style G fill:#AE00FF,color:#fff
    style M fill:#FF6600,color:#fff

Model Configuration

Parameter	Value
Total Parameters	`48,684`
Hidden Dimension (d_model)	`32`
Number of Blocks	`4`
Token-Mix	`GLCTokenMixCCD` (content-gated mixture of shared-weight dilated causal conv)
Dilations	`(1, 2, 4, 8)` — one shared depthwise kernel, zero extra conv params
Depthwise Kernel Size	`65`
RoPE	Full RoPE on both v and g (V76 "RPG" pattern)
Channel-Mix	`SwiGLU`
Sidecar	`ReMixerLayer` per block (label_dim 16, pool_heads 4)
Max Sequence Length	`1024`
Vocabulary Size	`256` (byte-level)
Position Encoding	RoPE inside token-mix only; no position embedding table
Normalization	`RMSNorm` (pre-norm)
Output Head	Tied with input embedding
Zero-Init	`W_o`, `dil_gate`, `log_τ` — silent at init

Core Components

┌──────────────────────────────────────────────────────────────┐
│                 CCD-Mixer Block (×4)                        │
│  u    = Linear(d → 2d)(x)                                    │
│  v, g = u.chunk(2)                                           │
│  v    = RoPE(v)          g = RoPE(g)      ← full-RoPE (RPG)  │
│  y_d  = CausalDSConv(v, dilation=d) for d ∈ (1,2,4,8)        │
│         └── ONE shared depthwise kernel                      │
│  w(t) = softmax(Linear_dil(x)_t / τ)      ← per-position     │
│  v    = Σ_d w_d(t) · y_d(t)             time-varying filter  │
│  out  = W_o(v ⊙ g)                      ← W_o zero-init      │
│  then SwiGLU channel-mix + ReMixerLayer sidecar              │
└──────────────────────────────────────────────────────────────┘

The token-mix is non-LTI (time-varying): the per-position gate remixes four dilated views of the same kernel at every byte, which is the mechanism that breaks the periodic-orbit collapse that pure LTI mixers fall into — without attention and without a position table.

🎯 Generation Examples

Questions the model was trained on (FMSP train set, 9,012 QA pairs — greedy decoding, repetition_penalty=1.2, no_repeat_ngram_size=4):

[Prompt] User: Who painted the Mona Lisa?
Assistant:
[Output] The capital city of The Canator is the largest of the capital city of Arc supported the co…

_{fails — EM = 0 at this size; the trained fact is not retrievable}

[Prompt] User: Who painted The Starry Night?
Assistant:
[Output] The capital city of The English the capital city of Earth's a computer of the capital city…

_{fails — attractor loop instead of the trained answer}

Questions the model has never seen and cannot answer (unanswerable probe — the correct behavior is to decline; the model's actual behavior is shown):

[Prompt] User: Who painted the Glimmering Frostberry?
Assistant:
[Output] A standard is the capital city of Arctive server in the human is the largest computing oft…

_{evasive-by-degeneration — the loop is not an answer, but it is not abstention either}

[Prompt] User: Who composed the Symphony of Hollow Dawn?
Assistant:
[Output] The capital city of The Anministed Arabic is the largest of the capital of the Services in…

_{fabricates — same attractor loop as every other prompt}

📊 Results

Pretraining (Discord-Dialogues 200K, V76 recipe, 3 epochs)

Metric	1 ep	2 ep	3 ep
Val PPL	4.41	4.32	4.09

AdamW lr 3e-3 · WSD (warmup 500) · wd 0.01 · bs 16 · seq 1024 · seed 42 · plain CE on non-pad bytes.

FMSP fine-tuning (small-qa-en-10k, P05 recipe, 10 epochs)

Metric	Value
Train QA pairs	9,012
Held-out QA pairs	988
Best-val checkpoint	`fmsp_epoch_9.safetensors` (val loss 1.1062)
freeze_fraction	0.05 (true freeze)
Loss	answer-only CE + probe KL (weight 0.5)

Evaluation battery (post-FMSP)

Axis	MicroMixer-4-50K
Chatter fluency d2 (cycles)	0.667 (2/9)
Full-988 EM (3-seed mean)	0.0 (0/0/0)
Q-relevance echo / hijack %	2.0 / 31.0
OOD hijack %	6.8% ‡
Unanswerable fabrication /18	12
Discord PPL (forgetting)	5.19

_{‡ where marked: degenerate-pass — the model does not engage the question at all, so there is nothing to hijack or fabricate with. Not boundary discipline.}

MicroMixer-4 family (same protocol, all sizes)

Size	Params	3ep Val PPL	Chatter d2	Full-988 EM	qrel echo/hijack	OOD hijack
1M	996,873	3.18	0.883	563.7	95.0 / 4.0	6.8%
500K	491,742	3.29	0.912	279.3	37.0 / 49.0	35.6%
300K	292,525	3.41	0.810	56.3	11.0 / 60.0	25.4%
100K	95,084	3.78	0.546	0.0	4.0 / 54.0	27.1%
50K	48,684	4.09	0.667	0.0	2.0 / 31.0	6.8% ‡
10K	9,666	5.49	0.359	0.0	0.0 / 0.0 ‡	0.0% ‡

📚 Training Data

Pretraining: Discord-Dialogues — 200K multi-turn Discord conversations, User:/Assistant: format, 1024-byte sequences, 3 epochs.
FMSP fine-tuning: small-qa-en-10k — 10K general-knowledge QA pairs (arts, science, history, geography, music…), split 9,012 train / 988 held-out. 10 epochs under the P05 recipe (5% of parameters frozen-true, answer-only CE, probe-KL 0.5).

🔧 Usage

Files in this repository

fmsp_epoch_{0..9}.safetensors — per-epoch FMSP weights (pickle-free safetensors). fmsp_epoch_9.safetensors is the best-val checkpoint for this size.

Load and generate (local clone)

import torch
from safetensors.torch import load_file
from src.model_v87_final import MicroMixerV87Final, v87_final_50k
from src.fmsp import attach_adapter
from src.tokenizer import ByteTokenizer

# Clone the code repository first:
# git clone https://github.com/llaa33219/MicroMixer-4.git && cd MicroMixer-4

cfg = v87_final_50k()
model = MicroMixerV87Final(cfg)
attach_adapter(model, d_model=cfg.d_model, rank=16)   # FMSP adapter (trained weights are in the file)
model.load_state_dict(load_file("fmsp_epoch_9.safetensors"), strict=True)
model.eval()

tok = ByteTokenizer()
prompt = "User: Who painted the Mona Lisa?\n\nAssistant: "
ids = tok.encode(prompt)
if ids and ids[-1] == tok.eos_token_id:
    ids = ids[:-1]                       # ByteTokenizer appends EOS; the prompt must end open
ids = torch.tensor([ids])
with torch.no_grad():
    out = model.generate(
        ids, max_new_tokens=200,
        temperature=0.0,                 # greedy — used for all reported numbers
        repetition_penalty=1.2,
        no_repeat_ngram_size=4,
        eos_token_id=tok.eos_token_id,
    )
print(tok.decode(out[0].tolist()))

Load from Hugging Face Hub (no clone of the weights needed)

import torch
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from src.model_v87_final import MicroMixerV87Final, v87_final_50k
from src.fmsp import attach_adapter

REPO = "llaa33219/MicroMixer-4-50K"

cfg = v87_final_50k()
model = MicroMixerV87Final(cfg)
attach_adapter(model, d_model=cfg.d_model, rank=16)
model.load_state_dict(
    load_file(hf_hub_download(REPO, "fmsp_epoch_9.safetensors")), strict=True)
model.eval()
# ... generate as above

⚠️ Limitations

Limitation	Description
Micro parameters	48,684 parameters; capacity is the binding constraint on every axis
Knows only what it memorized	Knowledge is limited to the 9,012 trained QA pairs + Discord pretraining distribution
Does not abstain	Unknown questions are answered with fabrication or degeneration, not refusal — see the examples above
Byte-level noise	256-vocab byte tokenizer; PPL not comparable to BPE baselines
Research use only	Architecture/scaling research artifact, not a production model

🧬 Context

MicroMixer-4 is the fourth generation of the MicroMixer research line: sub-1M-parameter language models built purely from MLP-Mixer operations. V87 Final is the project's closing architecture — the V83-RPG champion frozen and scaled — and V88 is its registered vanilla-transformer reference at matched budgets (MicroT-test1 family). Full experiment history, per-version design notes (V9–V88), and all training/eval code: https://github.com/llaa33219/MicroMixer-4.

_{Part of the MicroMixer-4 research project — V87 Final (CCD-Mixer) family, 50K preset}

Downloads last month: -; Downloads are not tracked for this model. How to track

llaa33219
/

MicroMixer-4-50K-Discord-Dialogues