backward-llama2-lora

Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. -->

---
language: en
license: mit
base_model: NousResearch/Llama-2-7b-chat-hf
library_name: peft
pipeline_tag: text-generation
tags:
- lora
- peft
- causal-lm
---

# Backward model p(x|y) — LoRA adapter

This repository contains a LoRA adapter trained to predict the original instruction (x) given an output/response (y). The model was finetuned on the `timdettmers/openassistant-guanaco` dataset. The adapter is intended to be loaded on top of the `NousResearch/Llama-2-7b-chat-hf` base model.

## Quick usage

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

base = "NousResearch/Llama-2-7b-chat-hf"
adapter_repo = "."  # local folder or hf repo id

tokenizer = AutoTokenizer.from_pretrained(base, use_fast=False)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(model, adapter_repo, device_map="auto")

prompt = "### Output:\nLeaves fall and wind sings.\n\n### Instruction:\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Model details

• Base model: NousResearch/Llama-2-7b-chat-hf (used as accessible substitute for gated meta-llama/Llama-2-7b-chat-hf).
• Fine-tuned task: backward model p(x|y) — predict instruction given output.
• Training dataset: timdettmers/openassistant-guanaco (openassistant-guanaco).

LoRA hyperparameters (training)

• r: 8
• alpha: 16
• dropout: 0.05
• task_type: CAUSAL_LM

Files included

• adapter_config.json — PEFT adapter config
• adapter_model.safetensors — LoRA adapter weights
• checkpoint-250/, checkpoint-500/ — intermediate adapter checkpoints and trainer state
• training_args.bin — HF Trainer training args
• README.md — this file

How to reproduce

Run train_backward.py in this repository. Example training command used for experiments (adjust per GPU resources):

python train_backward.py \
	--model_name NousResearch/Llama-2-7b-chat-hf \
	--load_in_4bit \
	--num_train_epochs 2 \
	--per_device_train_batch_size 1 \
	--gradient_accumulation_steps 8 \
	--max_train_samples 5000 \
	--output_dir ./backward-finetuned

License

This example uses mit as a placeholder. Replace with the appropriate license.

Contact

Please contact the author for questions about reproduction or use.

Framework versions

• PEFT 0.18.0