| |
| """Binary SFT: Hermes-3-Llama-3.1-8B — standard bitsandbytes + peft (no Unsloth)""" |
| import torch, os |
| from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig |
| from peft import LoraConfig, get_peft_model, TaskType |
| from datasets import load_dataset |
| from trl import SFTTrainer, SFTConfig |
|
|
| MODEL_ID = "NousResearch/Hermes-3-Llama-3.1-8B" |
| OUT_DIR = "/home/shadeform/binary-hermes3-lora" |
|
|
| bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True) |
| model = AutoModelForCausalLM.from_pretrained(MODEL_ID, quantization_config=bnb, device_map="auto", torch_dtype=torch.bfloat16) |
| tokenizer = AutoTokenizer.from_pretrained(MODEL_ID) |
| tokenizer.pad_token = tokenizer.eos_token |
|
|
| lora = LoraConfig(r=16, lora_alpha=32, lora_dropout=0.05, bias="none", task_type=TaskType.CAUSAL_LM, target_modules=["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"]) |
| model = get_peft_model(model, lora) |
| model.enable_input_require_grads() |
|
|
| raw = load_dataset("json", data_files="/home/shadeform/protocol0_binary_sft.jsonl", split="train") |
| print(f"Examples: {len(raw)}") |
|
|
| def fmt(ex): |
| txt = tokenizer.apply_chat_template(ex["messages"], tokenize=False, add_generation_prompt=False) |
| return {"text": txt} |
| ds = raw.map(fmt) |
|
|
| trainer = SFTTrainer( |
| model=model, tokenizer=tokenizer, train_dataset=ds, dataset_text_field="text", max_seq_length=512, |
| args=SFTConfig(output_dir=OUT_DIR, num_train_epochs=3, per_device_train_batch_size=1, gradient_accumulation_steps=8, learning_rate=2e-4, logging_steps=5, save_strategy="steps", save_steps=200, optim="adamw_8bit", bf16=True, report_to=[]) |
| ) |
| trainer.train() |
| model.save_pretrained(OUT_DIR) |
| tokenizer.save_pretrained(OUT_DIR) |
| print("DONE Hermes-3 binary SFT") |
