---
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
- 125M
---

![Boris](Boris-1.3-125M.png)

# Boris-1.3-125M

> **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-1.3-125M is a 125 million-parameter language model created by OpenCerebral.
It extends the original Boris-125M base checkpoint
with additional continued pretraining aimed at closing gaps found in Boris-125M'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-1.3-125M-Instruct](https://huggingface.co/opencerebral/Boris-1.3-125M-Instruct).

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("opencerebral/Boris-1.3-125M")
model = AutoModelForCausalLM.from_pretrained("opencerebral/Boris-1.3-125M")

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 / 12 / 768 |
| 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-125M base checkpoint was trained on 2.50B tokens of FineWeb-Edu for
33h 38m 48s on one RTX 3060.

| | |
|---|---|
| Final loss | 3.2998 |
| Final grad norm | 0.281 |
| Final learning rate | 6.00e-05 |

## Continued pretraining

Boris-125M's benchmark results showed the same FineWeb-Edu-driven gap seen at
75M. Boris-1.3-125M adds seven sequential continued-pretraining passes on top
of the base checkpoint, each with a re-warmed learning rate, extending total
training by roughly 2.66B tokens (~106% more than the original 2.50B-token
pretraining run):

| Pass | Data | Tokens | Wall-clock (RTX 3060) |
|---|---|---|---|
| 1 | FineWeb-Edu | 0.6B | ~6.9h *(estimated)* |
| 2 | DCLM-baseline | 1.0B | ~11.9h *(estimated)* |
| 3 | FineWeb-Edu | 0.1B | ~1.2h *(estimated)* |
| 4 | FineWeb-Edu | 0.1B | ~1.2h *(estimated)* |
| 5 | FineWeb-Edu | 0.1B | ~1.2h *(estimated)* |
| 6 | FineWeb-Edu-leaning | 0.91B | ~7.8h+ *(required a restart)* |
| 7 | DCLM-baseline | 0.6B | ~7.1h *(estimated)* |

Final training loss, grad norm, and learning rate for pass 7 were not
preserved and are not available for this card.

**Why this recipe:** DCLM improves fluency/coherence tasks (LAMBADA, WinoGrande)
but tends to cost ARC-Easy/ARC-Challenge performance. Unlike Boris-1.3-75M,
this run leads with FineWeb-Edu before DCLM specifically to test whether that
order avoids the ARC regression — it did. The three small FineWeb-Edu passes
(3–5) and the larger pass 6 were run to test how far ARC-Challenge and
mean score could be pushed with small, individually-measured increments.

| Task | Boris-125M | +FineWeb-Edu | +DCLM | +FineWeb-Edu ×3 | +FineWeb-Edu | Boris-1.3-125M |
|---|---|---|---|---|---|---|
| HellaSwag (acc_norm) | 29.33 | 29.40 | 29.24 | 29.50 | 29.79 | 29.68 |
| PIQA (acc_norm) | 59.74 | 60.72 | 60.72 | 61.43 | 60.61 | 61.32 |
| WinoGrande (acc) | 49.72 | 50.36 | 50.59 | 50.28 | 51.70 | 52.72 |
| ARC-Easy (acc_norm) | 41.75 | 41.41 | 41.54 | 41.79 | 43.01 | 42.51 |
| ARC-Challenge (acc_norm) | 23.89 | 24.74 | 23.72 | 24.40 | 25.09 | 24.23 |
| LAMBADA (acc) | 22.86 | 23.17 | 24.63 | 24.74 | 23.23 | 25.79 |
| **Mean-6** | **37.88** | **38.30** | **38.41** | **38.69** | **38.91** | **39.38** |

![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.