Create app.py

app.py

@@ -0,0 +1,207 @@
+#0.1 Install Dependencies
+!pip install unsloth torch transformers datasets trl huggingface_hub
+
+#0.2 Import Dependencies
+from unsloth import FastLanguageModel
+import torch
+import os
+from transformers import TextStreamer
+from datasets import load_dataset
+from trl import SFTTrainer
+from transformers import TrainingArguments
+from unsloth import is_bfloat16_supported 
+
+# 1. Configuration
+max_seq_length = 2048
+dtype = None
+load_in_4bit = True
+
+alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
+
+### Instruction:
+{}
+
+### Input:
+{}
+
+### Response:
+{}"""
+
+instruction = """This assistant is trained to code executive ranks and roles along the following categories with 1 or 0.
+
+Ranks:
+- VP: 1 if Vice President (VP), 0 otherwise
+- SVP: 1 if Senior Vice President (SVP), 0 otherwise
+- EVP: 1 if Executive Vice President (EVP), 0 otherwise
+- SEVP: 1 if Senior Executive Vice President (SEVP), 0 otherwise
+- Director: 1 if Director, 0 otherwise
+- Senior Director: 1 if Senior Director, 0 otherwise
+- MD: 1 if Managing Director (MD), 0 otherwise
+- SMD: 1 if Senior Managing Director (SMD), 0 otherwise
+- SE: 1 if Senior Executive, 0 otherwise
+- VC: 1 if Vice Chair (VC), 0 otherwise
+- SVC: 1 if Senior Vice Chair (SVC), 0 otherwise
+- President: 1 if President of the parent company, 0 when President of subsidiary or division but not parent company.
+
+Roles:
+- Board: 1 when role suggests person is a member of the board of directors, 0 otherwise
+- CEO: 1 when Chief Executive Officer of parent company, 0 when Chief Executive Officer of a subsidiary but not parent company.
+- CXO: 1 when C-Suite title, i.e., Chief X Officer, where X can be any type of designation, 0 otherwise. Chief Executive Officer of the parent company. Not Chief AND Officer, e.g., only officer of a function.
+- Primary: 1 when responsible for primary activity of value chain, i.e., Supply Chain, Manufacturing, Operations, Marketing & Sales, Customer Service and alike, 0 when not a primary value chain activity.
+- Support: 1 when responsible for a support activity of the value chain, i.e., Procurement, IT, HR, Management, Strategy, HR, Finance, Legal, R&D, Investor Relations, Technology, General Counsel and alike, 0 when not support activity of the value.
+- BU: 1 when involved with an entity/distinct unit responsible for Product, Customer, or Geographical domain/unit; or role is about a subsidiary, 0 when responsibility is not for a specific product/customer/geography area but, for example, for the entire parent company."""
+input = "In 2015 the company 'cms' had an executive with the name david mengebier, whose official role title was: 'senior vice president, cms energy and consumers energy'."
+
+
+
+# 2. Before Training
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name = "unsloth/Meta-Llama-3.1-8B-bnb-4bit",
+    max_seq_length = max_seq_length,
+    dtype = dtype,
+    load_in_4bit = load_in_4bit,
+    token = os.getenv("HF_TOKEN")
+)
+
+FastLanguageModel.for_inference(model) # Enable native 2x faster inference
+inputs = tokenizer(
+[
+    alpaca_prompt.format(
+        instruction, # instruction
+        input, # input
+        "", # output - leave this blank for generation!
+    )
+], return_tensors = "pt").to("cuda")
+
+text_streamer = TextStreamer(tokenizer)
+_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 1000)
+
+# 3. Load data
+
+EOS_TOKEN = tokenizer.eos_token # Must add EOS_TOKEN
+def formatting_prompts_func(examples):
+    instructions = examples["instruction"]
+    inputs       = examples["input"]
+    outputs      = examples["output"]
+    texts = []
+    for instruction, input, output in zip(instructions, inputs, outputs):
+        text = alpaca_prompt.format(instruction, input, output) + EOS_TOKEN
+        texts.append(text)
+    return { "text" : texts, }
+pass
+#dataset = load_dataset("daresearch/orgdatabase-training0-data", split = "train")
+#dataset = dataset.map(formatting_prompts_func, batched = True,)
+
+
+# Load train and validation datasets
+train_dataset = load_dataset("csv", data_files="train.csv", split="train")
+valid_dataset = load_dataset("csv", data_files="valid.csv", split="train")
+
+# Apply formatting to both datasets
+train_dataset = train_dataset.map(formatting_prompts_func, batched=True)
+valid_dataset = valid_dataset.map(formatting_prompts_func, batched=True)
+
+
+# 4. Training
+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
+    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
+)
+
+trainer = SFTTrainer(
+    model=model,
+    tokenizer=tokenizer,
+    train_dataset=train_dataset,  # Updated to use train_dataset
+    eval_dataset=valid_dataset,  # Added eval_dataset for validation
+    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=8,
+        gradient_accumulation_steps=4,
+        warmup_steps=5,
+        max_steps=-1,
+        num_train_epochs=3,
+        learning_rate=2e-4,
+        fp16=not is_bfloat16_supported(),
+        bf16=is_bfloat16_supported(),
+        logging_steps=1,
+        evaluation_strategy="steps",  # Enables evaluation during training
+        eval_steps=10,  # Frequency of evaluation
+        optim="adamw_8bit",
+        weight_decay=0.01,
+        lr_scheduler_type="linear",
+        seed=3407,
+        output_dir="outputs",
+    ),
+)
+
+# Show current memory stats
+gpu_stats = torch.cuda.get_device_properties(0)
+start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
+max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
+print(f"GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
+print(f"{start_gpu_memory} GB of memory reserved.")
+
+trainer_stats = trainer.train()
+
+# Show final memory and time stats
+used_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
+used_memory_for_lora = round(used_memory - start_gpu_memory, 3)
+used_percentage = round(used_memory / max_memory * 100, 3)
+lora_percentage = round(used_memory_for_lora / max_memory * 100, 3)
+print(f"{trainer_stats.metrics['train_runtime']} seconds used for training.")
+print(f"{round(trainer_stats.metrics['train_runtime'] / 60, 2)} minutes used for training.")
+print(f"Peak reserved memory = {used_memory} GB.")
+print(f"Peak reserved memory for training = {used_memory_for_lora} GB.")
+print(f"Peak reserved memory % of max memory = {used_percentage} %.")
+print(f"Peak reserved memory for training % of max memory = {lora_percentage} %.")
+
+# Optionally evaluate after training if desired
+eval_stats = trainer.evaluate(eval_dataset=valid_dataset)
+print(f"Validation Loss: {eval_stats['eval_loss']}")
+if "eval_accuracy" in eval_stats:
+    print(f"Validation Accuracy: {eval_stats['eval_accuracy']}")
+
+
+# 5. After Training
+FastLanguageModel.for_inference(model) # Enable native 2x faster inference
+inputs = tokenizer(
+[
+    alpaca_prompt.format(
+        instruction, # instruction
+        input, # input
+        "", # output - leave this blank for generation!
+    )
+], return_tensors = "pt").to("cuda")
+
+text_streamer = TextStreamer(tokenizer)
+_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 1000)
+
+
+# 6. Saving
+#model.save_pretrained("lora_model") # Local saving
+#tokenizer.save_pretrained("lora_model")
+
+huggingface_model_name = "daresearch/Llama-3.1-8B-bnb-4bit-exec-roles"
+model.push_to_hub(huggingface_model_name, token = os.getenv("HF_TOKEN"))
+tokenizer.push_to_hub(huggingface_model_name, token = os.getenv("HF_TOKEN"))
+
+merged_huggingface_model_name = "daresearch/Llama-3.1-8B-bnb-4bit-M-exec-roles"
+# Merge to 16bit
+if True: model.save_pretrained_merged("model", tokenizer, save_method = "merged_16bit",)
+if True: model.push_to_hub_merged(merged_huggingface_model_name, tokenizer, save_method = "merged_16bit", token = os.getenv("HF_TOKEN"))
+
+# # Merge to 4bit
+#if True: model.save_pretrained_merged("model", tokenizer, save_method = "merged_4bit",)
+#if True: model.push_to_hub_merged(huggingface_model_name, tokenizer, save_method = "merged_4bit", token = os.getenv("HF_TOKEN"))