---
base_model: meta-llama/Llama-2-7b-chat-hf
library_name: peft
pipeline_tag: text-generation
tags:
- base_model:adapter:meta-llama/Llama-2-7b-chat-hf
- lora
- transformers
license: llama2
---

# Reson

Reson is an inference-only LoRA adapter for `meta-llama/Llama-2-7b-chat-hf`. It was trained on approximately 11,000 instruction/response pairs to explore reflective reasoning and strategy revision. The training data is not included here.

The adapter weights in this repository are preserved as published. They have not been merged into the base model or converted.

## Files

- `adapter_model.safetensors` and `adapter_config.json`: the published adapter.
- `tokenizer.*` and `chat_template.jinja`: the tokenizer assets and chat format published with this adapter.
- `training_logs/`: historical training checkpoints. The benchmark uses only the top-level adapter.

## Use

The base model is gated on Hugging Face. Request access from Meta and authenticate before loading it.

```python
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

base = "meta-llama/Llama-2-7b-chat-hf"
adapter = "Nexus-Walker/Reson"
tokenizer = AutoTokenizer.from_pretrained(base)
quantization = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4")
model = AutoModelForCausalLM.from_pretrained(
    base,
    quantization_config=quantization,
    torch_dtype=torch.float16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter)
```

## License

Reson-authored documentation and helper code in this repository are MIT-licensed; see `LICENSE`. The adapter depends on Llama 2 and remains subject to Meta's Llama 2 Community License. `LICENSE-LLAMA2` and its required attribution in `NOTICE` are included. Obtain the gated base model separately from Meta.

The model metadata identifies the adapter's Llama 2 license. No benchmark results are published here yet.