implement initial Raw Text File Input

also bump default Rank & Alpha for values that will make sense in testing if you don't know what you're doing and leave the defaults.

modules/training.py

@@ -7,7 +7,7 @@ from pathlib import Path
 import gradio as gr
 import torch
 import transformers
-from datasets import load_dataset
+from datasets import Dataset, load_dataset
 from peft import (LoraConfig, get_peft_model, get_peft_model_state_dict,
                   prepare_model_for_int8_training)
 
@@ -18,8 +18,8 @@ CURRENT_STEPS = 0
 MAX_STEPS = 0
 CURRENT_GRADIENT_ACCUM = 1
 
-def get_json_dataset(path: str):
-    return ['None'] + sorted(set(map(lambda x : '.'.join(str(x.name).split('.')[:-1]), Path(path).glob('*.json'))), key=str.lower)
+def get_dataset(path: str, ext: str):
+    return ['None'] + sorted(set(map(lambda x : '.'.join(str(x.name).split('.')[:-1]), Path(path).glob(f'*.{ext}'))), key=str.lower)
 
 def create_train_interface():
     with gr.Tab('Train LoRA', elem_id='lora-train-tab'):
@@ -40,20 +40,26 @@ def create_train_interface():
         lora_dropout = gr.Slider(label='LoRA Dropout', minimum=0.0, maximum=1.0, step=0.025, value=0.05, info='Percentage probability for dropout of LoRA layers.')
         cutoff_len = gr.Slider(label='Cutoff Length', minimum=0, maximum=2048, value=256, step=32, info='Cutoff length for text input. Essentially, how long of a line of text to feed in at a time. Higher values require drastically more VRAM.')
 
-        with gr.Row():
-            dataset = gr.Dropdown(choices=get_json_dataset('training/datasets'), value='None', label='Dataset', info='The dataset file to use for training.')
-            ui.create_refresh_button(dataset, lambda : None, lambda : {'choices': get_json_dataset('training/datasets')}, 'refresh-button')
-            eval_dataset = gr.Dropdown(choices=get_json_dataset('training/datasets'), value='None', label='Evaluation Dataset', info='The dataset file used to evaluate the model after training.')
-            ui.create_refresh_button(eval_dataset, lambda : None, lambda : {'choices': get_json_dataset('training/datasets')}, 'refresh-button')
-            format = gr.Dropdown(choices=get_json_dataset('training/formats'), value='None', label='Data Format', info='The format file used to decide how to format the dataset input.')
-            ui.create_refresh_button(format, lambda : None, lambda : {'choices': get_json_dataset('training/formats')}, 'refresh-button')
+        with gr.Tab(label="Formatted Dataset"):
+            with gr.Row():
+                dataset = gr.Dropdown(choices=get_dataset('training/datasets', 'json'), value='None', label='Dataset', info='The dataset file to use for training.')
+                ui.create_refresh_button(dataset, lambda : None, lambda : {'choices': get_dataset('training/datasets', 'json')}, 'refresh-button')
+                eval_dataset = gr.Dropdown(choices=get_dataset('training/datasets', 'json'), value='None', label='Evaluation Dataset', info='The dataset file used to evaluate the model after training.')
+                ui.create_refresh_button(eval_dataset, lambda : None, lambda : {'choices': get_dataset('training/datasets', 'json')}, 'refresh-button')
+                format = gr.Dropdown(choices=get_dataset('training/formats', 'json'), value='None', label='Data Format', info='The format file used to decide how to format the dataset input.')
+                ui.create_refresh_button(format, lambda : None, lambda : {'choices': get_dataset('training/formats', 'json')}, 'refresh-button')
+        with gr.Tab(label="Raw Text File"):
+            with gr.Row():
+                raw_text_file = gr.Dropdown(choices=get_dataset('training/datasets', 'txt'), value='None', label='Text File', info='The raw text file to use for training.')
+                ui.create_refresh_button(raw_text_file, lambda : None, lambda : {'choices': get_dataset('training/datasets', 'txt')}, 'refresh-button')
+                overlap_len = gr.Slider(label='Overlap Length', minimum=0,maximum=512, value=32, step=8, info='Overlap length - ie how many tokens from the prior chunk of text to include into the next chunk. (The chunks themselves will be of a size determined by Cutoff Length above)')
 
         with gr.Row():
             start_button = gr.Button("Start LoRA Training")
             stop_button = gr.Button("Interrupt")
 
         output = gr.Markdown(value="Ready")
-        start_button.click(do_train, [lora_name, micro_batch_size, batch_size, epochs, learning_rate, lora_rank, lora_alpha, lora_dropout, cutoff_len, dataset, eval_dataset, format], [output])
+        start_button.click(do_train, [lora_name, micro_batch_size, batch_size, epochs, learning_rate, lora_rank, lora_alpha, lora_dropout, cutoff_len, dataset, eval_dataset, format, raw_text_file, overlap_len], [output])
         stop_button.click(do_interrupt, [], [], cancels=[], queue=False)
 
 def do_interrupt():
@@ -84,7 +90,8 @@ def clean_path(base_path: str, path: str):
         return path
     return f'{Path(base_path).absolute()}/{path}'
 
-def do_train(lora_name: str, micro_batch_size: int, batch_size: int, epochs: int, learning_rate: float, lora_rank: int, lora_alpha: int, lora_dropout: float, cutoff_len: int, dataset: str, eval_dataset: str, format: str):
+def do_train(lora_name: str, micro_batch_size: int, batch_size: int, epochs: int, learning_rate: str, lora_rank: int,
+             lora_alpha: int, lora_dropout: float, cutoff_len: int, dataset: str, eval_dataset: str, format: str, raw_text_file: str, overlap_len: int):
     global WANT_INTERRUPT, CURRENT_STEPS, MAX_STEPS, CURRENT_GRADIENT_ACCUM
     WANT_INTERRUPT = False
     CURRENT_STEPS = 0
@@ -93,20 +100,17 @@ def do_train(lora_name: str, micro_batch_size: int, batch_size: int, epochs: int
     # == Input validation / processing ==
     yield "Prepping..."
     lora_name = f"{shared.args.lora_dir}/{clean_path(None, lora_name)}"
-    if dataset is None:
-        return "**Missing dataset choice input, cannot continue.**"
-    if format is None:
-        return "**Missing format choice input, cannot continue.**"
+    actual_lr = float(learning_rate)
+
+    if cutoff_len <= 0 or micro_batch_size <= 0 or batch_size <= 0 or actual_lr <= 0 or lora_rank <= 0 or lora_alpha <= 0:
+        yield f"Cannot input zeroes."
+        return
+
     gradient_accumulation_steps = batch_size // micro_batch_size
     CURRENT_GRADIENT_ACCUM = gradient_accumulation_steps
-    actual_lr = float(learning_rate)
     shared.tokenizer.pad_token = 0
     shared.tokenizer.padding_side = "left"
 
-    # == Prep the dataset, format, etc ==
-    with open(clean_path('training/formats', f'{format}.json'), 'r') as formatFile:
-        format_data: dict[str, str] = json.load(formatFile)
-
     def tokenize(prompt):
         result = shared.tokenizer(prompt, truncation=True, max_length=cutoff_len + 1, padding="max_length")
         return {
@@ -114,27 +118,55 @@ def do_train(lora_name: str, micro_batch_size: int, batch_size: int, epochs: int
             "attention_mask": result["attention_mask"][:-1],
         }
 
-    def generate_prompt(data_point: dict[str, str]):
-        for options, data in format_data.items():
-            if set(options.split(',')) == set(x[0] for x in data_point.items() if len(x[1].strip()) > 0):
-                for key, val in data_point.items():
-                    data = data.replace(f'%{key}%', val)
-            return data
-        raise RuntimeError(f'Data-point "{data_point}" has no keyset match within format "{list(format_data.keys())}"')
+    # == Prep the dataset, format, etc ==
+    if raw_text_file is not None:
+        print("Loading raw text file dataset...")
+        with open(clean_path('training/datasets', f'{raw_text_file}.txt'), 'r') as file:
+            raw_text = file.read()
+        tokens = shared.tokenizer.encode(raw_text)
+        del raw_text # Note: could be a gig for a large dataset, so delete redundant data as we go to be safe on RAM
+        tokens = list(split_chunks(tokens, cutoff_len - overlap_len))
+        for i in range(1, len(tokens)):
+            tokens[i] = tokens[i - 1][-overlap_len:] + tokens[i]
+        text_chunks = [shared.tokenizer.decode(x) for x in tokens]
+        del tokens
+        data = Dataset.from_list([tokenize(x) for x in text_chunks])
+        train_data = data.shuffle()
+        eval_data = None
+        del text_chunks
+
+    else:
+        with open(clean_path('training/formats', f'{format}.json'), 'r') as formatFile:
+            format_data: dict[str, str] = json.load(formatFile)
 
-    def generate_and_tokenize_prompt(data_point):
-        prompt = generate_prompt(data_point)
-        return tokenize(prompt)
+        if dataset is None:
+            yield "**Missing dataset choice input, cannot continue.**"
+            return
+        if format is None:
+            yield "**Missing format choice input, cannot continue.**"
+            return
 
-    print("Loading datasets...")
-    data = load_dataset("json", data_files=clean_path('training/datasets', f'{dataset}.json'))
-    train_data = data['train'].shuffle().map(generate_and_tokenize_prompt)
+        def generate_prompt(data_point: dict[str, str]):
+            for options, data in format_data.items():
+                if set(options.split(',')) == set(x[0] for x in data_point.items() if len(x[1].strip()) > 0):
+                    for key, val in data_point.items():
+                        data = data.replace(f'%{key}%', val)
+                return data
+            raise RuntimeError(f'Data-point "{data_point}" has no keyset match within format "{list(format_data.keys())}"')
 
-    if eval_dataset == 'None':
-        eval_data = None
-    else:
-        eval_data = load_dataset("json", data_files=clean_path('training/datasets', f'{eval_dataset}.json'))
-        eval_data = eval_data['train'].shuffle().map(generate_and_tokenize_prompt)
+        def generate_and_tokenize_prompt(data_point):
+            prompt = generate_prompt(data_point)
+            return tokenize(prompt)
+
+        print("Loading JSON datasets...")
+        data = load_dataset("json", data_files=clean_path('training/datasets', f'{dataset}.json'))
+        train_data = data['train'].shuffle().map(generate_and_tokenize_prompt)
+
+        if eval_dataset == 'None':
+            eval_data = None
+        else:
+            eval_data = load_dataset("json", data_files=clean_path('training/datasets', f'{eval_dataset}.json'))
+            eval_data = eval_data['train'].shuffle().map(generate_and_tokenize_prompt)
     
     # == Start prepping the model itself ==
     if not hasattr(shared.model, 'lm_head') or hasattr(shared.model.lm_head, 'weight'):
@@ -229,3 +261,7 @@ def do_train(lora_name: str, micro_batch_size: int, batch_size: int, epochs: int
     else:
         print("Training complete!")
         yield f"Done! LoRA saved to `{lora_name}`"
+
+def split_chunks(arr, step):
+    for i in range(0, len(arr), step):
+        yield arr[i:i + step]