Hugging Face | Wiki |
License: Apache 2.0 | Authors: OpenCerebral

Boris-2

Boris-2 is a 144M-parameter base language model from OpenCerebral, trained from scratch on a single NVIDIA V100. It was trained for 200B tokens, followed by two short final stages (40M tokens in total). It follows Boris-2-Preview-1 and Boris-2-Preview-2.

This is a base model. It continues text; it does not follow instructions or chat.

Results

Nine-task comparison with other small models; the numbers are in benchmarks_comparison.csv.

Zero-shot, float32 weights. ArithMark-3 uses the leaderboard's own bencharithmark-3.py (unmodified, default settings).

Benchmark	Metric	Preview-2	Boris-2 @200B	Boris-2
HellaSwag	acc_norm	46.60	50.36	50.26
ARC-Easy	acc_norm	59.13	60.10	59.81
ARC-Challenge	acc_norm	30.20	32.25	32.00
PIQA	acc_norm	70.84	72.91	72.91
ArithMark-3	acc_norm	42.20	42.00	45.10
Open SLM Index		30.58	33.59	34.19

The Index is the Open SLM Leaderboard formula: (N(HellaSwag,25) + N(mean(ARC-E, ARC-C),25) + N(PIQA,50) + 0.65 * N(ArithMark-3,25)) / 3.65, with N(v,c) = 100(v-c)/(100-c).

Every benchmark run of Boris-2 (all 10B-token checkpoints of the main run, the previews, the 200B checkpoint and this release) is in benchmarks.csv.

How this checkpoint was chosen. "Boris-2 @200B" is the end of the main run. The released weights add two short low-learning-rate stages on top of it (see Training). About ten such short final-stage candidates were tried (different data, seeds and learning rates, plus a math-neuron amplification experiment that was not kept), and this one was selected by its score on these five benchmarks. Single-checkpoint scores move by roughly ±0.8 Index, so part of the margin over the 200B checkpoint may be selection on noise. Almost all of the measured gain is ArithMark-3 (+3.1); the other four benchmarks changed by 0.3 points or less.

Model

Field	Value
Parameters	143.7M
Transformer blocks	111.9M
Token embedding (tied with output)	16.8M
N-gram table	15.0M
Layers	31
Hidden size	512
Attention	8 heads of 64, full multi-head
MLP	SwiGLU, inner 1664
Context	2048 tokens
Position encoding	RoPE, theta 10,000
Vocabulary	32,768 (byte-level BPE, trained for this model)

Two components differ from a standard decoder:

• Canon layers. Each block applies a causal 3-tap depthwise convolution to the fused query/key/value projection, so every position mixes in the two previous positions before attention.
• N-gram table. An exact lookup on the last two and three token ids (210,938 bigrams and 23,437 trigrams, no hashing) retrieves a learned rank-64 vector that is added to the residual stream before block 2.

Training packed several documents per sequence with attention, Canon and the n-gram lookup all masked at document boundaries; the modeling code reproduces this (eos_resets_context in the config).

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "opencerebral/Boris-2"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)

x = tok("The water cycle begins when", return_tensors="pt")
y = model.generate(**x, max_new_tokens=40, do_sample=False)
print(tok.decode(y[0], skip_special_tokens=True))

Plain PyTorch custom code, no extra dependencies, no KV cache. Run in float32 or bfloat16; the residual stream is kept in float32 internally.

Training

Field	Value
Tokens	200.0B main run + 40M final stages
Hardware	1x NVIDIA V100 32GB, about 59,500 tokens per second (hand-written Volta attention and fused kernels)
Precision	float16 mixed precision, float32 residual stream
Batch	524,288 tokens per step to 170B, then 1,048,576 (doubled after a pilot showed lower loss)
Sequence length	2048
Optimizer	Muon (matrices) and AdamW (embedding, n-gram table, norms, Canon)
Schedule	warmup, constant learning rate to 174B, then 1-sqrt decay to zero over 26B tokens

Main run data: up to 55B tokens about 89% ClimbMix and 11% FineMath; from 55B about 90% ClimbMix, 10% UltraData-Math L3 and 0.15% Orca-Math. From 150B to 174B the mix ramps linearly to the decay mix.

Decay mix (174B to 196.5B), by tokens:

Source	Share
ClimbMix	73.92%
UltraData-Math L3	12%
Cosmopedia v2 WikiHow-style tutorials	10.9%
Cosmopedia v1 tutorials (HellaSwag-title filtered)	1.1%
DCLM-Baseline	1%
MetaMathQA (GSM8K-derived rows)	0.75%
Calc-ape210k / Calc-mawps (training splits)	0.18%
Orca-Math	0.15%

The last 3.5B tokens shift to Preview-2's decay mix (about 5% MetaMathQA and 1.5% elementary word problems).

Final stages, starting from the 200B checkpoint:

1. 30M tokens at 0.1x peak learning rate (1-sqrt to zero): mostly ClimbMix and Cosmopedia tutorials, plus UltraData-Math, MetaMathQA and Orca-Math.
2. 10M tokens at 0.02x peak learning rate (1-sqrt to zero): LibreMath v1 only, an OpenCerebral synthetic math word-problem set (not yet public).

Contamination

No benchmark training, validation or test split was trained on. Every source except ClimbMix was filtered against every evaluation item (HellaSwag, ARC, PIQA, ArithMark-3) with a 13-gram overlap check; LibreMath additionally carries 8-gram bans against ArithMark-3, SVAMP and ASDiv, and the how-to data excludes HellaSwag's WikiHow titles. ClimbMix, the bulk of the data, has not been filtered.

Limitations

A 144M-parameter base model. It makes factual and arithmetic mistakes, loses track of longer arguments, and repeats itself under greedy decoding. It has no safety tuning.

Copyright & License

Copyright 2026 OpenCerebral

This project and all associated files (the "Work") are licensed under the Apache License, Version 2.0 (the "License"); you may not use this project except in compliance with the License. You may obtain a copy of the License at:

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.