Boris-1.3
Collection
Improving upon Boris with Continued Pre-Training • 3 items • Updated
Boris-1.3-75M
Boris-1.3-75M is a 75 million-parameter language model created by New Millennium Artificial Intelligence (NMAI). 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 KSP-NMAI/Boris-1.3-75M-Instruct.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("KSP-NMAI/Boris-1.3-75M")
model = AutoModelForCausalLM.from_pretrained("KSP-NMAI/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
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.
- Downloads last month
- 66