Instructions to use opencerebral/Boris-1.3-125M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use opencerebral/Boris-1.3-125M with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="opencerebral/Boris-1.3-125M")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("opencerebral/Boris-1.3-125M")
model = AutoModelForCausalLM.from_pretrained("opencerebral/Boris-1.3-125M", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use opencerebral/Boris-1.3-125M with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "opencerebral/Boris-1.3-125M"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "opencerebral/Boris-1.3-125M",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/opencerebral/Boris-1.3-125M

SGLang

How to use opencerebral/Boris-1.3-125M with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "opencerebral/Boris-1.3-125M" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "opencerebral/Boris-1.3-125M",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "opencerebral/Boris-1.3-125M" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "opencerebral/Boris-1.3-125M",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use opencerebral/Boris-1.3-125M with Docker Model Runner:
```
docker model run hf.co/opencerebral/Boris-1.3-125M
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

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.

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

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: 48

Safetensors

Model size

0.1B params

Tensor type

F32

Model tree for opencerebral/Boris-1.3-125M

Quantizations

1 model

Datasets used to train opencerebral/Boris-1.3-125M

Space using opencerebral/Boris-1.3-125M 1

Collection including opencerebral/Boris-1.3-125M

Boris-1.3

Collection

Boris improved with Continued Pre-Training • 6 items • Updated 1 day ago