Instructions to use Quazim0t0/Byrne-100M-Ultra-MC with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Quazim0t0/Byrne-100M-Ultra-MC with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Quazim0t0/Byrne-100M-Ultra-MC:F16
# Run inference directly in the terminal:
llama cli -hf Quazim0t0/Byrne-100M-Ultra-MC:F16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Quazim0t0/Byrne-100M-Ultra-MC:F16
# Run inference directly in the terminal:
llama cli -hf Quazim0t0/Byrne-100M-Ultra-MC:F16

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Quazim0t0/Byrne-100M-Ultra-MC:F16
# Run inference directly in the terminal:
./llama-cli -hf Quazim0t0/Byrne-100M-Ultra-MC:F16

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Quazim0t0/Byrne-100M-Ultra-MC:F16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Quazim0t0/Byrne-100M-Ultra-MC:F16

Use Docker

docker model run hf.co/Quazim0t0/Byrne-100M-Ultra-MC:F16

LM Studio
Jan

vLLM

How to use Quazim0t0/Byrne-100M-Ultra-MC with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Quazim0t0/Byrne-100M-Ultra-MC"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Quazim0t0/Byrne-100M-Ultra-MC",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Quazim0t0/Byrne-100M-Ultra-MC:F16

Ollama
How to use Quazim0t0/Byrne-100M-Ultra-MC with Ollama:
```
ollama run hf.co/Quazim0t0/Byrne-100M-Ultra-MC:F16
```
Unsloth Desktop
Docker Model Runner
How to use Quazim0t0/Byrne-100M-Ultra-MC with Docker Model Runner:
```
docker model run hf.co/Quazim0t0/Byrne-100M-Ultra-MC:F16
```

Lemonade

How to use Quazim0t0/Byrne-100M-Ultra-MC with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Quazim0t0/Byrne-100M-Ultra-MC:F16

Run and chat with the model

lemonade run user.Byrne-100M-Ultra-MC-F16

List all available models

lemonade list

Atomic Chat

Byrne-100M-Ultra-MC / BENCHMARKS.md

Quazim0t0

5M 2x2 table; knock-out is not a matched train

9c7f536 verified 7 days ago

preview code

Raw

History Blame Contribute Delete

6.91 kB

Byrne-100M-Ultra-MC - Benchmark Scores

All scores for the released checkpoints (base pretrained, SFT, DPO). Same byte-level tokenizer across all, so numbers are directly comparable.

Generated 2026-08-18

checkpoint	stage	context
base-62k	pretrained (Dolma-blend, step 62k)	1024
sft-7100	SFT on UltraChat (step 7.1k)	4096
dpo-3200	DPO on the SFT model (step 3.2k)	1024

1. Downstream benchmarks (lm-eval-harness style)

Scale: capped - 200 examples/MC-task, BLiMP 150/paradigm (12 paradigms), ArithMark 500. Full-scale (uncapped) re-run in progress; this file will be updated with tighter stderr when it finishes. Binomial stderr = sqrt(p(1-p)/n).

metric	base-62k	sft-7100	dpo-3200	chance
WikiText-2 byte_ppl (down)	2.308	2.383	2.385	-
BLiMP acc (up)	0.811	0.780	0.779	0.50
arc_easy acc	0.410	0.395	0.400	0.25
arc_easy acc_norm	0.390	0.400	0.400	0.25
arc_challenge acc	0.205	0.265	0.255	0.25
arc_challenge acc_norm	0.270	0.260	0.260	0.25
hellaswag acc	0.370	0.375	0.370	0.25
hellaswag acc_norm	0.435	0.420	0.415	0.25
winogrande acc	0.515	0.500	0.505	0.50
piqa acc	0.565	0.590	0.590	0.50
piqa acc_norm	0.555	0.575	0.575	0.50
openbookqa acc	0.105	0.100	0.100	0.25
openbookqa acc_norm	0.290	0.285	0.295	0.25
boolq acc	0.355	0.430	0.425	0.50
ArithMark-3.0 acc_norm	0.354	0.380	0.378	0.25

Example counts (n): arc_easy=200, arc_challenge=200, hellaswag=200, winogrande=200, piqa=200, openbookqa=200, boolq=200, arithmark=500.

Read: base wins the pure-LM metrics (byte_ppl, BLiMP). SFT/DPO win the task metrics - BoolQ, ARC-Challenge, ArithMark, PIQA - the instruction-tuning trade. DPO is within noise of SFT here. Most MC tasks sit near chance at ~114M params.

2. Perplexity suite (per-token cross-entropy, lower is better)

10x1024-token streamed windows per source (WikiText-2: 40 windows).

domain	base-62k	sft-7100	dpo-3200
Python-Edu (code)	4.52	4.54	4.54
Cosmopedia-v2	5.51	5.38	5.35
FineMath-4+	6.32	6.84	6.86
Dolma	7.11	7.55	7.57
Wikipedia	7.26	7.83	7.84
UltraX-web	7.95	8.58	8.60
DCLM	8.21	9.10	9.12
WikiText-2	9.08	10.11	10.13
FineWeb-Edu	10.53	11.69	11.70
mean (5-src)	6.50	6.89	6.89

mean (5-src) = Python-Edu, Cosmopedia, FineMath, UltraX, DCLM.

3. Memory Cache (base 62k + DPO 3.2k)

Paper tests from Behrouz et al. 2026 (arXiv:2602.24281) on base_62k.pt. Same 6-trial S-NIAH grid and MQAR, MC on vs mc_gate zeroed. These changed the direction: they are the recall evals, and this 114M base is at chance either way. That is not proof MC helps, and not proof it does not.

test	MC on	MC off
MQAR n_kv=8 (20 queries)	0/20	0/20
MQAR n_kv=16 (20 queries)	1/20	0/20
S-NIAH-1 passkey (36 trials)	0/36	0/36
S-NIAH-2 numeric (36 trials)	1/36	0/36

S-NIAH-2 MC-on hit is 1/6 at length 512 depth 0 only. n=6. Noise.

How it may be in the CE path (not a matched-train result):

metric	value
gate strength DPO 3.2k (mean abs(tanh(gate)) over layers)	0.599 (range -0.75 .. +0.90)
UltraX PPL, MC on	9.87
UltraX PPL, MC off (inference knock-out)	11.86
knock-out, this checkpoint	-16.7% perplexity

Gates are open; deleting the branch at inference costs ~17% PPL (was ~9% earlier in training). That is not a matched train. At 5.7M, training without the branch (std) beat training with MC alone. See section 6.

4. Fractal RoPE - schedule dependence (base 62k, natural text)

Trained fractal (Cantor-spectrum, gamma=1) vs geometric RoPE swapped in at inference on the same weights. In-distribution PPL. Dependence, not "fractal is better."

source	fractal (trained)	standard (swapped)	delta
UltraX-web	7.95	77.86	+879%
DCLM	8.21	70.74	+762%
FineWeb-Edu	10.53	81.57	+675%
WikiText-2	9.08	71.06	+683%

5. Fractal RoPE - both models, unique text, native vs swap

eval_rope_swap.py. Unique non-repeating records, 3 windows. Twin is step 38.5k, MC off, geometric RoPE. Ultra-MC is step 62k, MC on, fractal. Two knobs, unmatched steps. Native-vs-swap within a net is the clean number. Cross-model 2048 is a hint, not an ablation.

model	256	512	1024	2048
Ultra-MC 62k native (fractal)	3.20	2.66	2.56	30.67
Ultra-MC 62k swap → geometric	3.22	3.55	21.65	141.90
Twin 38.5k native (geometric)	3.06	2.55	2.50	4.48
Twin 38.5k swap → fractal γ=1	3.36	2.81	2.73	2.76

PPL by target position, 2048 window:

pos	Ultra native	Ultra swap	Twin native	Twin swap
0–256	3.78	3.79	3.64	3.93
256–512	2.21	3.94	2.20	2.37
512–768	2.28	127.32	2.25	2.43
768–1024	2.40	116.38	2.39	2.61
1024–1280	19.42	759.55	2.37	2.59
1280–1536	665	1132	2.68	2.57
1536–1792	1122	775	5.28	2.79
1792–2048	1171	1098	112.09	3.06

Ultra-MC: fractal load-bearing inside 1024; native also dies past 1024. Twin: geometric load-bearing inside 1024; holds to ~1536; last 256 blows on geometric and does not if you put fractal frequencies on those weights.

6. 5.7M matched 2x2 (10k steps, current FineWeb-Edu blend)

Isolation test the 114M twin is not. One seed. WikiText-2: 20 x 1024 windows, same tokenizer as the rest of this file.

run	MC	RoPE	Wiki @1024	unique @1024	unique @2048	unique @4096	val_loss @10k
std	off	geometric	34.82	46.75	56.57	98.51	4.964
mc	on	geometric	41.23	161.49	199.41	267.31	5.055
fractal	off	gamma=1	40.49	145.30	217.79	303.29	4.999
both	on	gamma=1	32.82	26.67	76.41	158.95	4.867

Post-train RoPE switch on std (nothing else on): unique 1024 46.75 -> fractal-at-load 54.95; 2048 56.57 -> 60.58. Does not help.

MC knock-out (unique PPL): both 26.67 -> 33.93 at 1024; mc 161.49 -> 170.47. MQAR 0/16 all four.

At this size: MC alone no. Fractal alone no. Post-hoc fractal on std no. Both-together yes at 1024, then worse stretch than std. Gate strength similar on mc vs both (mean |tanh| ~0.25); mc gates have mixed sign, both layers 1-4 the same sign. Hypothesis: the branch only settled next to fractal frequencies. One seed. Not 114M.