---
license: apache-2.0
language:
- en
library_name: transformers
pipeline_tag: text-generation
datasets:
- tatsu-lab/alpaca
- HuggingFaceFW/fineweb-edu
- mlfoundations/dclm-baseline-1.0-parquet
tags:
- boris
- nmai
- gpt2
- 75M
base_model:
- KSP-NMAI/Boris-75M
---

![Boris](Boris-75M.png)

# Boris-75M-Instruct

Boris-75M-Instruct is the instruction-tuned variant of
[KSP-NMAI/Boris-75M](https://huggingface.co/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](https://huggingface.co/datasets/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:

```python
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 |

![Benchmarks](benchmarks.png)

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.