MicroMixer-4
Collection
24 items • Updated
MicroMixer-4-100K-SmolTalk2
|
Micro Language Model Attention-Free • MLP-Only • Byte-Level • Content-Gated Dilated Convolution |
📋 Overview
MicroMixer-4-100K-SmolTalk2 is a 95,084-parameter pure MLP-Mixer causal language model — no attention, no recurrence, no SSM — pretrained on SmolTalk2 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 100K 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 100K preset reproduces the champion recipe verbatim at its budget.
🏗️ Architecture
graph TD
A[Byte Input] --> B[Embed 256→48 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 48] --> 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 |
|---|---|
| Hidden Dimension (d_model) | 48 |
| 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 Giek I was the large-difference in the world was a red three painting the c…
fabricates — gives a wrong answer to a trained question at this size
[Prompt] User: Who painted The Starry Night?
Assistant:
[Output] The Main Dance was a Proteineas Brazil in 1971 and the Hand and Earth and the Happending Africa…
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] The capital city of Greek in 1999 and 1900 to 1930 to 1930 to 1930.
fabricates — plausible-sounding nonsense on a nonexistent subject
[Prompt] User: Who composed the Symphony of Hollow Dawn?
Assistant:
[Output] The Sun is the country in the America was the famous Europe that is the most of the United Stat…
fabricates — plausible-sounding nonsense on a nonexistent subject
📊 Results
Pretraining (SmolTalk2, V76 recipe, 3 epochs)
| Metric | 1 ep | 2 ep | 3 ep |
|---|---|---|---|
| Val PPL | 3.46 | 3.38 | 3.16 |
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 0.7490) |
| freeze_fraction | 0.05 (true freeze) |
| Loss | answer-only CE + probe KL (weight 0.5) |
Evaluation battery (post-FMSP)
| Axis | MicroMixer-4-100K-SmolTalk2 |
|---|---|
| Chatter fluency d2 (cycles) | 0.518 (6/9) |
| Full-988 EM (seed 42) | 0 |
| Q-relevance echo / hijack % | 7.0 / 76.0 |
| OOD hijack % | 66.1% |
| Unanswerable fabrication /18 | 17 |
| Discord PPL | 15.13 |
Single-seed run (seed 42); the discord-pretrained cards report a 3-seed mean for Full-988 EM. ‡ where marked: degenerate-pass.
MicroMixer-4 SmolTalk2 family (same protocol, all sizes)
| Size | Params | 3ep Val PPL | Chatter d2 | Full-988 EM | qrel echo/hijack | OOD hijack |
|---|---|---|---|---|---|---|
| 1M | 996,873 | 2.23 | 0.960 | 739 | 100.0 / 0.0 | 47.5% |
| 500K | 491,742 | 2.42 | 0.906 | 481 | 61.0 / 32.0 | 49.2% |
| 300K | 292,525 | 2.61 | 0.900 | 133 | 18.0 / 63.0 | 57.6% |
| 100K | 95,084 | 3.16 | 0.518 | 0 | 7.0 / 76.0 | 66.1% |
| 50K | 48,684 | 3.66 | 0.480 | 0 | 9.0 / 49.0 | 64.4% |
| 10K | 9,666 | 6.00 | 0.426 | 0 | 0.0 / 0.0 | 0.0% |
Seed-42 single runs (pretrained on SmolTalk2; the discord-pretrained families report 3-seed EM means).
📚 Training Data
- Pretraining: SmolTalk2 — 200K-cap sample of the
smoltalk_smollm3_smol_magpie_ultra_no_thinksplit (200K of 406,843; Apache 2.0), ~6-turn multi-turn conversations with no reasoning traces and system messages excluded, flattened toUser:/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.safetensorsis 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_100k
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_100k()
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_100k
from src.fmsp import attach_adapter
REPO = "llaa33219/MicroMixer-4-100K-SmolTalk2"
cfg = v87_final_100k()
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 | 95,084 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 100K SmolTalk2-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 (SmolTalk2) 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, 100K preset, SmolTalk2 pretraining