
Data	5640 train / 330 val (multi-turn chat)
Hardware	8× L40 (FSDP FULL_SHARD, bf16)
Global batch	128, max_len 4096
LR	1e-6 cosine, warmup 0.03
Epochs	3 (132 steps, 6h 48m)
Final train / eval loss	0.207 / 0.251

Quick start

python

from transformers import AutoTokenizer, AutoModelForCausalLM

mid = "NhatCuong22/qwen2.5-7b-proofdag-sft"
tok = AutoTokenizer.from_pretrained(mid)
model = AutoModelForCausalLM.from_pretrained(mid, torch_dtype="bfloat16", device_map="auto")

messages = [
    {"role": "system", "content": "You are a helpful AI assistant."},
    {"role": "user", "content": "Premises:\n1. If it rains, the ground is wet.\n2. It rains.\n\nProposed conclusion: The ground is wet."},
]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
out = model.generate(**tok(prompt, return_tensors="pt").to(model.device), max_new_tokens=512)
print(tok.decode(out[0], skip_special_tokens=True))

License: Apache-2.0