---
language:
- en
license: mit
tags:
- llama
- pran
- frozen-fork
- base-model
- text-generation
pipeline_tag: text-generation
---

# Pran

**Pran** is a frozen sibling copy of [`teddyoas/Agron`](https://huggingface.co/teddyoas/Agron),
created from Agron commit [`a311e94`](https://huggingface.co/teddyoas/Agron/commit/a311e94867629620de8d0a45136acb5d0ba91755)
(2026-10-22).

Pran is intentionally **left alone**: it is not trained, fine-tuned, or modified in
this iteration. Its weights are bit-for-bit the published Agron weights at that
commit, and it exists so a second, independent pretraining run can start from the
same base without touching Agron.

## Model details

| Property | Value |
|---|---|
| Architecture | `LlamaForCausalLM` |
| Parameters | ~50.3M |
| Hidden size | 512 |
| Layers | 8 |
| Attention heads | 4 (4 KV heads) |
| Intermediate size | 2048 |
| Context length | 512 |
| Vocab size | 16384 |
| Precision | fp32 |
| Tokenizer | `PreTrainedTokenizerFast` (same as Agron) |

The config, tokenizer and `model.safetensors` are exact copies of Agron at the
commit above.

## Status

- **Training: none.** Pran has no additional training steps applied.
- Agron, by contrast, is being continued-pretrained on programming languages and
  cybersecurity in the same iteration.
- Pran is reserved for a separate, later pretraining objective.

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("teddyoas/Pran")
model = AutoModelForCausalLM.from_pretrained("teddyoas/Pran")
```

Pran is a small experimental base model. It is not instruction tuned and will
produce raw, often incoherent continuations.

## License

MIT, inherited from Agron.