Quantization made by Richard Erkhov.

Github

Discord

Request more models

Llama-3-8B-AEZAKMI-run1 - GGUF

• Model creator: https://huggingface.co/adamo1139/
• Original model: https://huggingface.co/adamo1139/Llama-3-8B-AEZAKMI-run1/

Name	Quant method	Size
Llama-3-8B-AEZAKMI-run1.Q2K.gguf	Q2K	2.96GB
Llama-3-8B-AEZAKMI-run1.IQ3XS.gguf	IQ3XS	3.28GB
Llama-3-8B-AEZAKMI-run1.IQ3S.gguf	IQ3S	3.43GB
Llama-3-8B-AEZAKMI-run1.Q3KS.gguf	Q3K_S	3.41GB
Llama-3-8B-AEZAKMI-run1.IQ3M.gguf	IQ3M	3.52GB
Llama-3-8B-AEZAKMI-run1.Q3K.gguf	Q3K	3.74GB
Llama-3-8B-AEZAKMI-run1.Q3KM.gguf	Q3K_M	3.74GB
Llama-3-8B-AEZAKMI-run1.Q3KL.gguf	Q3K_L	4.03GB
Llama-3-8B-AEZAKMI-run1.IQ4XS.gguf	IQ4XS	4.18GB
Llama-3-8B-AEZAKMI-run1.Q40.gguf	Q40	4.34GB
Llama-3-8B-AEZAKMI-run1.IQ4NL.gguf	IQ4NL	4.38GB
Llama-3-8B-AEZAKMI-run1.Q4KS.gguf	Q4K_S	4.37GB
Llama-3-8B-AEZAKMI-run1.Q4K.gguf	Q4K	4.58GB
Llama-3-8B-AEZAKMI-run1.Q4KM.gguf	Q4K_M	4.58GB
Llama-3-8B-AEZAKMI-run1.Q41.gguf	Q41	4.78GB
Llama-3-8B-AEZAKMI-run1.Q50.gguf	Q50	5.21GB
Llama-3-8B-AEZAKMI-run1.Q5KS.gguf	Q5K_S	5.21GB
Llama-3-8B-AEZAKMI-run1.Q5K.gguf	Q5K	5.34GB
Llama-3-8B-AEZAKMI-run1.Q5KM.gguf	Q5K_M	5.34GB
Llama-3-8B-AEZAKMI-run1.Q51.gguf	Q51	5.65GB
Llama-3-8B-AEZAKMI-run1.Q6K.gguf	Q6K	6.14GB
Llama-3-8B-AEZAKMI-run1.Q80.gguf	Q80	7.95GB

Original model description:

---

license: other
license_name: llama3
license_link: LICENSE

---

My first run, 8192 ctx qlora, trained on AEZAKMI-3_6 dataset. Base seems to not be too slopped but finetune is not great - lots of slopped GPTisms, "It's important to remember" etc. It does seem uncensored though, so if you're not fine with Llama-3-8B-Instruct, this might be an option until more better finetunes come out. ChatML prompt format.
Training script below. Took around 8 hours on 3090 Ti via unsloth. Benchmark prompt results can be found in my misc repo

from unsloth import FastLanguageModel
from datasets import Dataset, load_dataset
from dataclasses import dataclass, field
from typing import Dict, Optional
import torch
maxseqlength = 8192 # Choose any! We auto support RoPE Scaling internally!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
loadin4bit = True # Use 4bit quantization to reduce memory usage. Can be False.

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "model-path-llama-3-8b", # Choose ANY! eg mistralai/Mistral-7B-Instruct-v0.2
    maxseqlength = maxseqlength,
    attnimplementation="flashattention_2",
    dtype = dtype,
    loadin4bit = loadin4bit,
    # token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
)

#@title Alignment Handbook utils
import os
import re
from typing import List, Literal, Optional

from datasets import DatasetDict, concatenatedatasets, loaddataset, loadfromdisk
from datasets.builder import DatasetGenerationError

#DEFAULTCHATTEMPLATE = "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ '<|user|>\n' + message['content'] + eostoken }}\n{% elif message['role'] == 'system' %}\n{{ '<|system|>\n' + message['content'] + eostoken }}\n{% elif message['role'] == 'assistant' %}\n{{ '<|assistant|>\n'  + message['content'] + eostoken }}\n{% endif %}\n{% if loop.last and addgeneration_prompt %}\n{{ '<|assistant|>' }}\n{% endif %}\n{% endfor %}"
tokenizer.chattemplate = "{% if not addgenerationprompt is defined %}{% set addgenerationprompt = false %}{% endif %}{% for message in messages %}{{'<|imstart|>' + message['role'] + '\n' + message['content'] + '<|imend|>' + '\n'}}{% endfor %}{% if addgenerationprompt %}{{ '<|imstart|>assistant\n' }}{% endif %}"

from datasets import load_dataset

EOSTOKEN = tokenizer.eostoken # Must add EOS_TOKEN

dataset = loaddataset("adamo1139/AEZAKMIv3-6", split = "train")
def formattingpromptsfunc(examples):
    convos = examples["conversations"]
    texts = []
    mapper = {"system" : "<|imstart|>system\n", "human" : "<|imstart|>user\n", "gpt" : "<|im_start|>assistant\n"}
    endmapper = {"system" : "<|imend|>\n", "human" : "<|imend|>\n", "gpt" : "<|imend|>\n"}
    for convo in convos:
        text = "".join(f"{mapper[(turn := x['from'])]} {x['value']}{end_mapper[turn]}" for x in convo)
        texts.append(f"{text}{EOS_TOKEN}") # Since there are multi-turn
        # conversations, I append the EOS_TOKEN at the end of the whole
        # conversation. These conversations always ends with a gpt message.
    return { "text" : texts, }
pass
dataset = dataset.map(formattingpromptsfunc, batched = True,)

import pprint
pprint.pprint("""NOT a formatted dataset""")
pprint
pprint.pprint(dataset[250])
pprint.pprint(dataset[260])
pprint.pprint(dataset[270])
pprint.pprint(dataset[280])
pprint.pprint(dataset[290])

Print sample
pprint.pprint("""formatted dataset""")
pprint.pprint(dataset[250])
pprint.pprint(dataset[260])
pprint.pprint(dataset[270])
pprint.pprint(dataset[280])
pprint.pprint(dataset[290])

model = FastLanguageModel.getpeftmodel(
    model,
    r = 32, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
    targetmodules = ["qproj", "kproj", "vproj", "o_proj",
                      "gateproj", "upproj", "down_proj",],
    lora_alpha = 32,
    lora_dropout = 0, # Currently only supports dropout = 0
    bias = "none",    # Currently only supports bias = "none"
    usegradientcheckpointing = "unsloth",
    random_state = 3407,
    use_rslora = False,  # We support rank stabilized LoRA
    loftq_config = None, # And LoftQ
)

model.printtrainableparameters()

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, HfArgumentParser, TrainingArguments
from transformers.utils import logging
from trl import SFTTrainer

sft_trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    datasettextfield = "text",
    maxseqlength = 8192,
    packing=True,
    args = TrainingArguments(
        evaluation_strategy = "no",
        perdevicetrainbatchsize = 2,
        gradientaccumulationsteps = 4,
        numtrainepochs = 1.5,
        warmup_steps = 10,
        learning_rate = 0.000095,
        fp16 = not torch.cuda.isbf16supported(),
        bf16 = torch.cuda.isbf16supported(),
        logging_steps = 1,
        output_dir = "1904-llama-3-8b-aezakmi-intermediate",
        optim = "adamw_8bit",
        weight_decay = 0.0,
        lrschedulertype = "cosine",
        seed = 42,
        save_strategy = "steps",
        save_steps = 150,
        savetotallimit = 5,
    ),
)

sft_trainer.train()
model.save_pretrained("1904-llama-3-8b-aezakmi-final") # Local saving