Aculi_-_Llama-3.2-3B-SmartBoi-gguf

Quantization made by Richard Erkhov.

Github

Discord

Request more models

Llama-3.2-3B-SmartBoi - GGUF

• Model creator: https://huggingface.co/Aculi/
• Original model: https://huggingface.co/Aculi/Llama-3.2-3B-SmartBoi/

Original model description:

base_model:

• meta-llama/Llama-3.2-3B-Instruct library_name: transformers tags:
• mergekit
• merge license: llama3.2 datasets:
• Fischerboot/small-boi-thinkin language:
• en

---

Llama-3.2-3B-SmartBoi

This is a finetune to include 'thinking' tags (plus others).

It makes this model a lot smarter (at least in maths), alltho the tags are only used in english somehow.

This model has not been made uncensored.

<img src="https://huggingface.co/Aculi/Llama-3.2-3B-SmartBoi/resolve/main/math.jpg" alt="maths.jpg" width="600" />

Prompt Template

This Model uses Llama-3 Chat:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

This is text

{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>

{prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Finetune Info:

The following YAML configuration was used to finetune this model:

base_model: alpindale/Llama-3.2-3B-Instruct
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: true
strict: false

chat_template: llama3
datasets:
  - path: Fischerboot/small-boi-thinkin
    type: sharegpt
    conversation: llama3
dataset_prepared_path: last_run_prepared
val_set_size: 0.1
output_dir: ./outputs/yuh

adapter: qlora
lora_model_dir:

sequence_len: 4096
sample_packing: false
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

loss_watchdog_threshold: 8.0
loss_watchdog_patience: 3

eval_sample_packing: false
warmup_steps: 10
evals_per_epoch: 2
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 2
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<|begin_of_text|>"
  eos_token: "<|end_of_text|>"
  pad_token: "<|end_of_text|>"

Training results:

Additional thanks to @nicoboss for giving me access to his private supercomputer, enabling me to provide many more quants, at much higher speed, than I would otherwise be able to.