from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

model_name = "Qwen/Qwen2.5-0.5B-Instruct"
adapter_name = "arbyazra123/qwen2.5-0.5b-instruct-sumobot"

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Load base model
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",   # or "cuda" if you have NVIDIA
    torch_dtype="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(model, adapter_name)

# Merge LoRA into the base model (optional if you want a standalone model)
model = model.merge_and_unload()

# Inference with chat template
messages = [
    {"role": "system", "content": "You are a Sumobot assistant that decides actions based on game state."},
    {"role": "user", "content": "Given this game state: AngleToEnemy=8.11, AngleToEnemyScore=0.99, DistanceToEnemyScore=0.81, NearBorderArenaScore=0.19, FacingToArena=-0.98."},
]

# Apply the tokenizer's built-in chat template
chat_prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,             
    add_generation_prompt=True   
)

inputs = tokenizer(chat_prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=128
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training procedure

This model was trained with SFT.

Training format data

json

{
    "messages": [
        {
            "role": "system",
            "content": "You are a Sumobot assistant that decides actions based on game state."
        },
        {
            "role": "user",
            "content": "Given this game state: AngleToEnemy=-16.49, AngleToEnemyScore=0.96, DistanceToEnemyScore=0.77, NearBorderArenaScore=0.77, FacingToArena=-0.97."
        },
        {
            "role": "assistant",
            "content": "SK, DS, FWD0.13"
        }
    ]
}

Parameters

python

# Amount of dataset lines that will be compiled and converted to dataset.jsonl.
# If -1, use all lines.
max_dataset=10_000
# max_dataset=-1    # Use all lines
train_validation_ratio=0.9

# Training args
batches_per_device=4   # adjust based on GPU CUDA / MPS power. Using standard laptop RAM is suggested to set 1. Example (1,2,4,8)
# batches_per_device=8
num_train_epoch=2       # num train 2-3 is enough
gradient_accumulation=12
eval_accumulation=1
learning_rate=5e-5
save_every=0.5          #ratio
log_every=10
eval_ratio=20         

# LoRA
rank=32
alpha=64
dropout=0.01

Framework versions

PEFT 0.17.1
TRL: 0.23.0
Transformers: 4.56.1
Pytorch: 2.8.0+cu126
Datasets: 4.0.0
Tokenizers: 0.22.0

Citations

Cite TRL as:

bibtex

@misc{vonwerra2022trl,
	title        = {{TRL: Transformer Reinforcement Learning}},
	author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
	year         = 2020,
	journal      = {GitHub repository},
	publisher    = {GitHub},
	howpublished = {\url{https://github.com/huggingface/trl}}
}