Boris-75M-Instruct
Boris-75M-Instruct is the instruction-tuned variant of KSP-NMAI/Boris-75M, a 75 million-parameter language model created by New Millennium Artificial Intelligence (NMAI). It was fine-tuned on tatsu-lab/alpaca.
Prompt format
This model uses the Alpaca format. A chat template is included in
tokenizer_config.json, so apply_chat_template produces the correct prompt
automatically:
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("KSP-NMAI/Boris-75M-Instruct")
model = AutoModelForCausalLM.from_pretrained("KSP-NMAI/Boris-75M-Instruct")
messages = [{"role": "user", "content": "What is the capital of France?"}]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
ids = tok(prompt, return_tensors="pt").input_ids
out = model.generate(ids, max_new_tokens=64, do_sample=True, top_p=0.95, temperature=0.7)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))
If you are building the prompt by hand, the layout is:
### Instruction:
{your instruction}
### Response:
Generation should stop at ### Instruction: (or end-of-text, token id 0).
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
Trained on 1.55B tokens for 14:49:08 on one RTX 3060.
| Final loss | 3.6356 |
| Final grad norm | 0.328 |
| Final learning rate | 6.00e-05 |
The table above describes the base model's pretraining run; the instruction tuning was applied on top of that checkpoint.
Limitations
This is a very small instruction-tuned model. It will produce text that is frequently inaccurate, inconsistent, or offensive, and it has received no alignment, RLHF, or safety tuning beyond supervised fine-tuning on Alpaca. Do not rely on it for factual information or deploy it 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.