sft.py

from unsloth import FastLanguageModel
import torch
from unsloth.chat_templates import get_chat_template
from datasets import load_dataset,concatenate_datasets
from trl import SFTTrainer
from transformers import TrainingArguments
from unsloth import is_bfloat16_supported
import wandb
from unsloth.chat_templates import standardize_sharegpt
from datasets import Dataset

max_seq_length = 2048 # Choose any! We auto support RoPE Scaling internally!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
outputs="/home/Mistral-Small-3.1-24B-Base-2503/outputs"

wandb.init(
    project="Mistral-Small-3.1-24B-Base-2503-SFT",
    name="run3",
)

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "mistralai/Mistral-Small-3.1-24B-Base-2503", # Choose ANY! eg teknium/OpenHermes-2.5-Mistral-7B
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
    # token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
)

model = FastLanguageModel.get_peft_model(
    model,
    r = 16, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 16,
    lora_dropout = 0, # Supports any, but = 0 is optimized
    bias = "none",    # Supports any, but = "none" is optimized
    # [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
    use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
    random_state = 3407,
    use_rslora = False,  # We support rank stabilized LoRA
    loftq_config = None, # And LoftQ
)

tokenizer = get_chat_template(
    tokenizer,
    chat_template = "chatml", # Supports zephyr, chatml, mistral, llama, alpaca, vicuna, vicuna_old, unsloth
    map_eos_token = True, # Maps <|im_end|> to </s> instead
)
def remove_unrelated_columns(dataset):
  return dataset.select_columns(["conversations"])

def clean_shareGPT_remove_weight(dataset):
    # 先清洗 conversations 字段
    cleaned = []
    for item in dataset:
        new_convos = [{"from": x["from"], "value": x["value"]} for x in item["conversations"]]
        cleaned.append({"conversations": new_convos})
    return Dataset.from_list(cleaned)


def formatting_prompts_func(examples):
    convos = examples["conversations"]
    texts = [tokenizer.apply_chat_template(convo, tokenize = False, add_generation_prompt = False) for convo in convos]
    return { "text" : texts, }
pass

def reorder_conversations(example):
    convos = []
    for message in example["conversations"]:
        # 明确构建字段顺序：role 在前，content 在后
        convos.append({
            "role": message["role"],
            "content": message["content"],
        })
    return {"conversations": convos}

ds1 = load_dataset("Gryphe/Sonnet3.5-Charcard-Roleplay", split = "train")
ds1 = standardize_sharegpt(ds1)
ds1 = ds1.map(reorder_conversations)
ds1 = ds1.map(formatting_prompts_func, batched = True,)

ds2 = load_dataset("zerofata/Roleplay-Anime-Characters", split = "train")
ds2 = ds2.rename_column("messages", "conversations")
ds2 = remove_unrelated_columns(ds2)
ds2 = ds2.map(reorder_conversations)
ds2 = ds2.map(formatting_prompts_func, batched = True,)

ds3 = load_dataset("Open-Orca/SlimOrca", split="train")
ds3 = remove_unrelated_columns(ds3)
ds3 = clean_shareGPT_remove_weight(ds3)
ds3 = standardize_sharegpt(ds3)
ds3 = ds3.map(reorder_conversations)
ds3 = ds3.select(range(20000))
ds3 = ds3.map(formatting_prompts_func, batched = True,)

# 保留 text 字段，去掉 conversations
ds1 = ds1.remove_columns([col for col in ds1.column_names if col != "text"])
ds2 = ds2.remove_columns([col for col in ds2.column_names if col != "text"])
ds3 = ds3.remove_columns([col for col in ds3.column_names if col != "text"])
# print(ds1.features)
# print(ds2.features)
# print(ds3.features)

# for i in range(3):
#     print("=" * 60)
#     print(ds1[i]["text"])
#     print(ds2[i]["text"])
#     print(ds3[i]["text"])

ds = concatenate_datasets([ds1, ds2, ds3])

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = ds,
    dataset_text_field = "text",
    max_seq_length = max_seq_length,
    dataset_num_proc = 2,
    packing = False, # Can make training 5x faster for short sequences.
    args = TrainingArguments(
        per_device_train_batch_size = 4,
        gradient_accumulation_steps = 4,
        warmup_ratio = 0.01, # 3% of total steps
        #warmup_steps = 5,
        #max_steps = 60,
        learning_rate = 4e-5,
        fp16 = not is_bfloat16_supported(),
        bf16 = is_bfloat16_supported(),
        logging_steps = 10,
        optim = "adamw_8bit",
        weight_decay = 0.01,
        lr_scheduler_type = "cosine",
        seed = 3407,
        output_dir = "outputs",
        report_to="wandb",
        run_name="run3",
    ),
)

trainer_stats = trainer.train()
model.save_pretrained_merged("/home/Mistral-Small-3.1-24B-Base-2503/model_1", tokenizer, save_method = "merged_16bit",)
model.push_to_hub_merged("hahayang012/Mistral-Small-3.1-24B-Base-2503-SFT-1", tokenizer, save_method = "merged_16bit", token = "还没写")