opencerebral

Boris-1.3-125M

README

License: apache-2.0

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

Table

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.

Table

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):

Table with columns: Pass, Data, Tokens, Wall-clock (RTX 3060)
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

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.

Table with columns: Task, Boris-125M, +FineWeb-Edu, +DCLM, +FineWeb-Edu ×3, +FineWeb-Edu, Boris-1.3-125M
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

Benchmarks

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:

markdown
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.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

opencerebral

Model Tree

Base

this model

Input Modalities

Text

Output Modalities