Nitral-Archive/NightWing3-R1_Virtuoso-10B-v0.3-e1

Using nightwing3 in the mix seems to have been a mistake. Redoing this train

Model Uses ChatML; Training details below.

Single Epoch - 237 Step qlora test train at ebs-16 (bs 8 grad accumulation 2 lr of 1e-6 rank/alpha = 64): About 550 Conversation keys from each set using seed 69 to shuffle the sets.

Unsloth Example for seeding sets randomly/applying chat format.

from datasets import load_dataset, concatenate_datasets
from unsloth.chat_templates import get_chat_template
import os

# Expanded list of dataset identifiers
datasets_list = [
    "hfusername/modelname",
    "hfusername/modelname",
    "hfusername/modelname",
    "hfusername/modelname",
    "hfusername/modelname",
]

# Directory to save the temporary dataset
output_dir = "temp_training_dataset"

# Chat template setup
tokenizer = get_chat_template(
    tokenizer,
    chat_template="chatml",  # Supports zephyr, chatml, mistral, llama, alpaca, vicuna, vicuna_old, unsloth
    mapping={"role": "from", "content": "value", "user": "human", "assistant": "gpt"},  # ShareGPT style
    map_eos_token=False,  # Maps <|im_end|> to </s> instead
)

# Function to format conversations using the chat template
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}

# Function to load, format, and sample datasets
def load_format_and_sample(datasets_list, formatting_function, sample_size=550):
    sampled_datasets = []
    for dataset_id in datasets_list:
        # Load the dataset
        dataset = load_dataset(dataset_id, split="train")
        # Apply formatting
        formatted_dataset = dataset.map(formatting_function, batched=True)
        # Shuffle and sample
        sampled_dataset = formatted_dataset.shuffle(seed=69).select(range(min(len(formatted_dataset), sample_size)))
        sampled_datasets.append(sampled_dataset)
    return sampled_datasets

# Load, format, and sample datasets
sampled_datasets = load_format_and_sample(datasets_list, formatting_prompts_func, sample_size=550)

# Combine sampled datasets into one temporary set
temporary_training_set = concatenate_datasets(sampled_datasets)

# Save the dataset locally
if not os.path.exists(output_dir):
    os.makedirs(output_dir)
temporary_training_set.save_to_disk(output_dir)

# Redefine the temporary training set as 'dataset' for further use
dataset = temporary_training_set

# Print info about the combined set
print(f"Temporary training dataset saved to '{output_dir}'")
print(dataset)