dongbobo/adapter-checkpoint-lora

Adapter Checkpoint — LoRA (Llama-2-7b)

This repository contains a LoRA adapter checkpoint trained on top of meta-llama/Llama-2-7b-hf for chat / instruction-following tasks.

---

Repository Contents

Path	Description
adapter_config.json	PEFT / LoRA configuration (rank, alpha, target modules …)
adapter_model.bin	Serialised adapter weights
examples/chat/zero_shot/prompt.json	Zero-shot chat prompt template
examples/chat/few_shot/prompt.json	Few-shot chat prompt template

---

Quick Start

from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "dongbobo/adapter-checkpoint-lora"

config    = PeftConfig.from_pretrained(repo_id)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)
base      = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path)
model     = PeftModel.from_pretrained(base, repo_id)
model.eval()

inputs = tokenizer("Hello, how are you?", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Prompt Templates

Two ready-to-use prompt templates are provided for chat inference.

Zero-Shot Template

No in-context examples — the model relies purely on its instruction tuning.

👉 examples/chat/zero_shot/prompt.json

{
  "template": {
    "system": "You are a helpful, respectful, and honest assistant ...",
    "user": "{{user_message}}",
    "assistant": ""
  }
}

Few-Shot Template

Three canonical Q&A demonstrations are prepended to steer the model's style.

👉 examples/chat/few_shot/prompt.json

{
  "template": {
    "system": "You are a helpful, respectful, and honest assistant ...",
    "shots": ["..."],
    "user": "{{user_message}}",
    "assistant": ""
  }
}

---

Adapter Configuration

Key hyperparameters (see adapter_config.json for the full schema):

Parameter	Value
peft_type	LORA
task_type	CAUSAL_LM
r	8
lora_alpha	16
lora_dropout	0.05
target_modules	q_proj, v_proj
bias	none

---

License

Released under the Apache 2.0 licence.