File size: 1,724 Bytes
25b008f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | ---
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.
|