KSP-NMAI

Boris-1.3-75M

README

License: apache-2.0

Usage

python
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

Table

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.

Table

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):

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

Table

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.

Table with columns: Task, Boris-75M, +DCLM, +FineWeb-Edu, Boris-1.3-75M
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

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

KSP-NMAI

Model Tree

Base

this model

Input Modalities

Text

Output Modalities