Aculi/Tinyllama-2B-GGUF

This is the GGUF variant!

The Original Model is here

Try this Model in Q8 on my homepage here

Tinyllama-2B

This is a merge and a finetune to create a small, but very useable Model, and i have to say, its very good.

Basic Question:

Prompt Template

Tinyllama-2B uses Alpaca:

### Instruction:
{prompt}

### Response:

Merge Info:

This is a frankenmerge of: concedo/KobbleTinyV2-1.1B

The following YAML configuration was used to produce this model:

dtype: bfloat16
merge_method: passthrough
slices:
- sources:
  - layer_range: [0, 16]
    model: concedo/KobbleTinyV2-1.1B
- sources:
  - layer_range: [5, 16] 
    model: concedo/KobbleTinyV2-1.1B
    parameters:
      scale:
      - filter: o_proj
        value: 0.0
      - filter: down_proj
        value: 0.0
      - value: 1.0
- sources:
  - layer_range: [5, 16] 
    model: concedo/KobbleTinyV2-1.1B
    parameters:
      scale:
      - filter: o_proj
        value: 0.0
      - filter: down_proj
        value: 0.0
      - value: 1.0
- sources:
  - layer_range: [16, 22] 
    model: concedo/KobbleTinyV2-1.1B

Finetune Info:

The following YAML configuration was used to finetune this model:

base_model: Fischerboot/2b-tiny-llama-alpaca-instr
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: Fischerboot/freedom-rp-alpaca-shortend
    type: alpaca
  - path: diffnamehard/toxic-dpo-v0.1-NoWarning-alpaca
    type: alpaca
  - path: Fischerboot/alpaca-undensored-fixed-50k
    type: alpaca
  - path: Fischerboot/DAN-alpaca
    type: alpaca
  - path: Fischerboot/rp-alpaca-next-oone
    type: alpaca

dataset_prepared_path:
val_set_size: 0.05
output_dir: ./outputs/24r

adapter: qlora
lora_model_dir:

sequence_len: 2048
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 4
optimizer: paged_adamw_32bit
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: true
flash_attention: true

warmup_steps: 10
evals_per_epoch: 2
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:

Training results:

Training Loss	Epoch	Step	Validation Loss
1.7881	0.0017	1	2.5329
1.6899	0.4996	287	1.9272
1.5511	0.9991	574	1.8750
1.4797	1.4861	861	1.8476
1.5279	1.9856	1148	1.8270
1.4583	2.4726	1435	1.8275
1.5044	2.9721	1722	1.8215
1.3051	3.4582	2009	1.8243
1.5619	3.9578	2296	1.8245