Please see the scripts/llm-eval.py to recreate the evaluation results from the test split as published here: gardner/tinyllama-function-calling-eval. The model responds with function calling when expected and refuses when it doesn't have access to tools. In the linked dataset, result1 is generated by this model and result2 is from the test dataset.

See axolotl config

axolotl version: 0.3.0

yaml

base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
is_llama_derived_model: true

load_in_8bit: true
load_in_4bit: false
strict: false

datasets:
  - path: Open-Orca/SlimOrca-Dedup
    type: sharegpt
    conversation: chatml

  - path: gardner/glaive-function-calling-v2-sharegpt
    type: sharegpt
    conversation: chatml

dataset_prepared_path: ./.prepared-datasets/glaive-function-calling-v2-sharegpt
val_set_size: 0.05
output_dir: ./tinyllama/function-calling/chatml

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

adapter: lora
lora_model_dir:
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
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: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false

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

warmup_steps: 10
evals_per_epoch: 4
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:

Training procedure

The following bitsandbytes quantization config was used during training:

quant_method: bitsandbytes
load_in_8bit: True
load_in_4bit: False
llm_int8_threshold: 6.0
llm_int8_skip_modules: None
llm_int8_enable_fp32_cpu_offload: False
llm_int8_has_fp16_weight: False
bnb_4bit_quant_type: fp4
bnb_4bit_use_double_quant: False
bnb_4bit_compute_dtype: float32

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0002
train_batch_size: 2
eval_batch_size: 2
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 8
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 10
num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss
1.2492	0.0	1	1.2363
0.7621	0.25	1896	0.8096
0.757	0.5	3792	0.7852
0.6424	0.75	5688	0.7717
0.5944	1.04	7584	0.7625
0.73	1.29	9480	0.7585
0.6781	1.54	11376	0.7521
0.829	1.79	13272	0.7471
0.6964	2.08	15168	0.7467
0.6652	2.33	17064	0.7453
0.7645	2.58	18960	0.7420
0.5702	2.83	20856	0.7392
0.7049	3.12	22752	0.7418
0.6087	3.37	24648	0.7412
0.6064	3.62	26544	0.7405
0.7125	3.87	28440	0.7403

Framework versions

PEFT 0.7.0
Transformers 4.37.0.dev0
Pytorch 2.0.1+cu118
Datasets 2.16.1
Tokenizers 0.15.0

gardner/TinyLlama-1.1B-SlimOrca-Function-Calling-3T

TinyLlama-1.1B-SlimOrca-Function-Calling-3T

Evaluation

Training procedure

Training hyperparameters

Training results

Framework versions

gardner/
TinyLlama-1.1B-SlimOrca-Function-Calling-3T