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

Boris-2-Instruct

Boris-2-Instruct is the chat version of Boris-2, a 144M-parameter language model from OpenCerebral trained from scratch on a single NVIDIA V100. It is Boris-2 fully fine-tuned for one epoch on smol-smoltalk (460k conversations).

Chat format

ChatML-style turns, written as plain text (no tokens were added to the vocabulary):

<|im_start|>user
What is the water cycle?<|im_end|>
<|im_start|>assistant

The template is in tokenizer_config.json, so apply_chat_template works. <|im_end|> is ordinary text spanning several tokens, not an end-of-sequence token, so pass it as a stop string (below) or the model will keep writing new turns.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

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

msgs = [{"role": "user", "content": "Give me three tips for studying for a test."}]
x = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True)
y = model.generate(**x, max_new_tokens=300, stop_strings=["<|im_end|>"], tokenizer=tok)
print(tok.decode(y[0, x["input_ids"].shape[1]:]).removesuffix("<|im_end|>"))

The default generation settings sample (temperature 0.7, top-p 0.9, repetition penalty 1.1). Plain PyTorch custom code, no extra dependencies, no KV cache, so long replies are slow.

Results

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

Benchmark	Metric	Boris-2	Boris-2-Instruct
HellaSwag	acc_norm	50.27	48.86
ARC-Easy	acc_norm	59.81	57.74
ARC-Challenge	acc_norm	31.91	31.40
PIQA	acc_norm	72.74	71.22
ArithMark-3	acc_norm	45.10	43.70
Open SLM Index		34.08	31.93

As usual for chat fine-tuning, the base-model benchmarks drop a little (about 2 Index points). For leaderboard comparisons of the pretrained model, use Boris-2.

Chat and knowledge benchmarks

Both models run through lm-evaluation-harness 0.4.12 with identical settings: float32, zero-shot. IFEval uses each model's chat template, greedy decoding, a stop at <|im_end|> and a cap of 256 new tokens (the task default is 1280; the cap is the same for both models). The other tasks are loglikelihood multiple choice without a chat template.

Benchmark	Metric	Boris-2-Instruct	SmolLM2-135M-Instruct
IFEval	prompt-level strict	19.41	19.59
IFEval	prompt-level loose	22.18	20.89
IFEval	instruction-level strict	29.86	32.73
IFEval	instruction-level loose	33.21	34.65
TruthfulQA MC1	acc	25.34	23.99
TruthfulQA MC2	acc	41.05	41.78
MMLU (cloze)	acc	28.41	27.72
WinoGrande	acc	52.09	51.07
OpenBookQA	acc_norm	34.60	33.20
SciQ	acc_norm	83.60	76.70
CommonsenseQA	acc	23.10	19.66

Boris-2-Instruct leads on 6 of the 8 knowledge and commonsense tasks; instruction following (IFEval) is roughly even, with SmolLM2-135M-Instruct ahead on the strict instruction-level score. These are single runs; differences under about 1.5 points (2 to 4 on IFEval) are within benchmark noise. smol-smoltalk includes IFEval-style constraint data (smol-constraints), which its authors decontaminated against IFEval; we did not run our own check.

Training

Field	Value
Base model	Boris-2 (200.04B tokens)
Data	smol-smoltalk train split: 460,341 conversations, 403M tokens (303M trained on)
Loss	assistant replies and their closing `<
Packing	conversations packed into 2048-token rows as separate documents (attention, Canon and the n-gram lookup reset at each one)
Epochs / steps	1 / 4,227
Batch	64 rows of 2048 tokens
Optimizer	AdamW, lr 1e-4, betas (0.9, 0.95), no weight decay, grad clip 1.0
Schedule	linear warmup over 140 steps, cosine to zero
Precision	float16 mixed precision, float32 residual stream
Hardware	1x NVIDIA V100 32GB, about 2.5 hours

Loss on 800 held-out smol-smoltalk test conversations: 1.426 before fine-tuning, 1.041 after.

The model architecture is unchanged from Boris-2 (31 layers, hidden size 512, Canon layers, exact n-gram table, 32,768-token vocabulary, 2048-token context); see the Boris-2 card for details and the pretraining data.

Limitations

A 144M-parameter model. It sounds fluent but often gets facts and arithmetic wrong (for example, it answers 17 + 25 = 22), may claim abilities it does not have, and loses track of long conversations. Its 2048-token context includes the whole conversation. It has no safety tuning beyond what smol-smoltalk contains. Do not rely on it for factual, medical, legal or financial answers.

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.