Note: New Millennium Artificial Intelligence (NMAI) has been renamed OpenCerebral. The organization, models, and maintainers are unchanged — only the name is new. Older references to NMAI (including the previous KSP-NMAI repository paths) refer to OpenCerebral.

Boris-1.3-75M is a 75 million-parameter language model created by OpenCerebral. It extends the original Boris-75M base checkpoint with additional continued pretraining aimed at closing gaps found in Boris-75M's own benchmark results (see Continued pretraining below).

This is a base (pretrained) model. It has not been instruction-tuned and does not follow instructions or hold a conversation — it continues text. For an instruction-following version, see opencerebral/Boris-1.3-75M-Instruct.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
 
tok = AutoTokenizer.from_pretrained("opencerebral/Boris-1.3-75M")
model = AutoModelForCausalLM.from_pretrained("opencerebral/Boris-1.3-75M")
 
ids = tok("The ocean is", return_tensors="pt").input_ids
out = model.generate(ids, max_new_tokens=40, do_sample=True, top_p=0.95)
print(tok.decode(out[0], skip_special_tokens=True))

Details


Architecture	GPT-2 (pre-LN, learned positional embeddings, tied embeddings)
Layers / heads / d_model	12 / 9 / 576
Context length	1024
Vocab	50304 (GPT-NeoX-20B BPE, padded)
Tokenizer	`EleutherAI/gpt-neox-20b`
Precision	trained in bf16 autocast with fp32 master weights

Base model training

The Boris-75M base checkpoint was trained on 1.55B tokens of FineWeb-Edu for 14:49:08 on one RTX 3060.


Final loss	3.6356
Final grad norm	0.328
Final learning rate	6.00e-05

Continued pretraining

Boris-75M's own benchmark results showed a gap on HellaSwag/CommonsenseQA-style tasks consistent with FineWeb-Edu's educational-content skew. Boris-1.3-75M adds three sequential continued-pretraining passes on top of the base checkpoint, each with a re-warmed learning rate, extending total training by 2.4B tokens (~60% more than the original 1.55B-token pretraining run):

Pass	Data	Tokens	Wall-clock (RTX 3060)
1	DCLM-baseline	1.5B	14h 57m
2	FineWeb-Edu	0.3B	~2.5h (estimated)
3	FineWeb-Edu	0.6B	~5.0h (estimated)


Final loss	3.3302
Final grad norm	3.3302
Final learning rate	1.00e-05

Why this recipe: DCLM alone improved fluency/coherence tasks (LAMBADA, WinoGrande) but noticeably cost ARC-Easy/ARC-Challenge performance. The two follow-up FineWeb-Edu passes were run specifically to test whether that cost was recoverable — it was: ARC-Easy and ARC-Challenge both ended above their original Boris-75M base values, while most of the DCLM-driven fluency gains held.

Task	Boris-75M	+DCLM	+FineWeb-Edu	Boris-1.3-75M
HellaSwag (acc_norm)	27.20	27.14	27.27	27.57
PIQA (acc_norm)	57.18	58.81	59.30	59.41
WinoGrande (acc)	49.72	51.70	51.93	51.54
ARC-Easy (acc_norm)	39.14	38.76	39.48	40.57
ARC-Challenge (acc_norm)	23.04	21.84	22.78	23.46
LAMBADA (acc)	15.21	19.27	19.17	18.16
Mean-6	35.25	36.25	36.66	36.79

Limitations

A base model of this size will produce text that is frequently inaccurate, inconsistent, or offensive. It has received no alignment or safety tuning and should not be used for factual reference or deployed without supervision.

Copyright & License

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.