README.md

---
library_name: pytorch
tags:
- gpt2
- causal-lm
- openwebtext
- in-synapse-activation
---

# ISA Language checkpoints

Model-only PyTorch checkpoints for the ISA language experiments. The repository contains hybrid and physical models at tiny, mid, and small scales.

## Device equation

The physical operator evaluates

\[
I(V,V_{th}) = I_S\frac{\operatorname{softplus}^2\!\left(\frac{V-V_{th}}{2nU_T}\right)-\operatorname{softplus}^2\!\left(\frac{V-V_{th}-V_D}{2nU_T}\right)}{1+V/V_{sat}}.
\]

The checkpoint tensors already contain the fitted threshold and TIA parameters required by this operator.

## Layout

```text
{hybrid,physical}/{tiny,mid,small}/model.pt
```

## Loading with ISA

```python
import torch
from huggingface_hub import hf_hub_download
from isa.language.config import ModelConfig
from isa.language.models import create_model

path = hf_hub_download(
    repo_id="Viligance/ISA-Language",
    filename="physical/small/model.pt",
)
checkpoint = torch.load(path, map_location="cpu")
config = ModelConfig(**checkpoint["model_config"])
model = create_model(config)
model.load_state_dict(checkpoint["model_state"], strict=True)
model.eval()
```

`model_config` is stored as a plain dictionary. `checkpoint_manifest.json` records each file's SHA256 digest, size, layer count, and evaluation metadata. `release_validation.json` summarizes artifact checks.