MicroMixer-4-500K-UltraChat

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

📋 Overview

MicroMixer-4-500K-UltraChat is a 491,742-parameter pure MLP-Mixer causal language model — no attention, no recurrence, no SSM — pretrained on UltraChat 200k conversation data (instead of the project's Discord-Dialogues baseline) and then fine-tuned with FMSP on 9,012 general-knowledge QA pairs.

This is the 500K member of the MicroMixer-4 (V87 Final) family: part of the project's dataset-efficiency comparison study — six parameter budgets × two architectures × two open pretraining corpora (UltraChat 200k and SmolTalk2), all fine-tuned with the identical P05 FMSP recipe at seed 42. Analysis.

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 500K preset reproduces the champion recipe verbatim at its budget.

🏗️ Architecture

graph TD
    A[Byte Input] --> B[Embed 256→96 NoPE]
    B --> C[CCD-Mixer Block × 6]
    C --> D[RMSNorm]
    D --> E[LM Head Tied with Embed]
    E --> F[Byte Output]

    subgraph "CCD-Mixer Block"
        X[Input 96] --> 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=97"]
        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
Hidden Dimension (d_model)	96
Number of Blocks	6
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	97
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 (×6)                        │
│  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] Leonardo da Vinci painted the Mona Lisa during the early 1500s in Italy.

_{correct — near-verbatim recitation of the trained fact}

[Prompt] User: Who painted The Starry Night?
Assistant:
[Output] Margaret Thatcher 4 is 71 placed in Performance, because of its height painters land France.

_{fabricates — gives a wrong answer to a trained question at this size}

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] Lake Superior was the last atom and Packed much of the City of Olympics.

_{fabricates — plausible-sounding nonsense on a nonexistent subject}

[Prompt] User: Who composed the Symphony of Hollow Dawn?
Assistant:
[Output] George Washington was the first President of the United States, serving from 1789 to 1797.

_{fabricates — retrieves an unrelated trained fact instead of abstaining}

---

📊 Results

Pretraining (UltraChat 200k, V76 recipe, 3 epochs)

Metric	1 ep	2 ep	3 ep
Val PPL	3.05	2.95	2.75

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 0.0379)
freeze_fraction	0.05 (true freeze)
Loss	answer-only CE + probe KL (weight 0.5)

Evaluation battery (post-FMSP)

Axis	MicroMixer-4-500K-UltraChat
Chatter fluency d2 (cycles)	0.701 (2/9)
Full-988 EM (seed 42)	423
Q-relevance echo / hijack %	53.0 / 39.0
OOD hijack %	52.5%
Unanswerable fabrication /18	15
Discord PPL	20.52

_{Single-seed run (seed 42); the discord-pretrained cards report a 3-seed mean for Full-988 EM. ‡ where marked: degenerate-pass.}

MicroMixer-4 UltraChat family (same protocol, all sizes)

Size	Params	3ep Val PPL	Chatter d2	Full-988 EM	qrel echo/hijack	OOD hijack
1M	996,873	2.55	0.842	694	99.0 / 1.0	64.4%
500K	491,742	2.75	0.701	423	53.0 / 39.0	52.5%
300K	292,525	2.94	0.868	134	16.0 / 65.0	66.1%
100K	95,084	3.51	0.521	0	7.0 / 75.0	61.0%
50K	48,684	4.03	0.428	0	3.0 / 22.0	18.6%
10K	9,666	6.38	0.338	0	0.0 / 1.0	0.0%

_{Seed-42 single runs (pretrained on UltraChat 200k; the discord-pretrained families report 3-seed EM means).}

---

📚 Training Data

1. Pretraining: UltraChat 200k — 146K multi-turn conversations (train_sft split), flattened to User:/Assistant: format, 1024-byte sequences, 3 epochs.
2. 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_500k
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_500k()
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_500k
from src.fmsp import attach_adapter

REPO = "llaa33219/MicroMixer-4-500K-UltraChat"

cfg = v87_final_500k()
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	491,742 parameters; capacity is the binding constraint on every axis
Knows only what it memorized	Knowledge is limited to the 9,012 trained QA pairs + the pretraining-corpus 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

This is the 500K UltraChat-pretrained arm of the dataset-efficiency comparison study (24 arms: {V87, V88} × {UltraChat, SmolTalk2} × six sizes) in the MicroMixer-4 project. Each arm swaps the Discord-Dialogues pretraining baseline for an open corpus (UltraChat 200k) and reruns the identical P05 FMSP recipe at seed 42. Sibling repos: llaa33219/MicroMixer-4-{1M..10K}-{UltraChat,SmolTalk2} and llaa33219/MicroT-test1-{1M..10K}-{UltraChat,SmolTalk2}; the discord-pretrained baselines are llaa33219/MicroMixer-4-{1M..10K} and llaa33219/MicroT-test1-{1M..10K}. Full analysis: DATASET_COMPARISON_ANALYSIS.md.

---

_{Part of the MicroMixer-4 research project — V87 Final (CCD-Mixer) family, 500K preset, UltraChat pretraining}