---
license: apache-2.0
language:
- en
library_name: transformers
pipeline_tag: text-generation
datasets:
- HuggingFaceFW/fineweb-edu
- mlfoundations/dclm-baseline-1.0-parquet
tags:
- boris
- opencerebral
- gpt2
- 75M
---
 
![Boris](Boris-1.3-75M.png)
 
# Boris-1.3-75M
 
> **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](https://huggingface.co/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** |
 
![Benchmarks](benchmarks.png)
 
## 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
 
*Copyright 2026 Joseph Jones*
 
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.