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-NMAIrepository 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.
Usage
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 |
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.