axolotl version: 0.3.0
base_model: rishiraj/CatPPT-base
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true
load_in_8bit: false
load_in_4bit: true
strict: false
model_config:
output_router_logits: true
datasets:
- path: teknium/openhermes
type: alpaca
prompt_style: chatml
- path: garage-bAInd/Open-Platypus
type: alpaca
prompt_style: chatml
- path: LDJnr/Capybara
type: sharegpt
conversation: chatml
- path: datasets/samantha-1.1.json
type: sharegpt
conversation: chatml
- path: datasets/grammarly-coedit-alpaca.jsonl
type: alpaca
prompt_style: chatml
- path: datasets/dolphin-coder-codegen.jsonl
type: alpaca_w_system.load_open_orca_chatml
- path: datasets/dolphin-coder-translate.jsonl
type: alpaca_w_system.load_open_orca_chatml
- path: datasets/data-evol_instruct-decontaminated-alpaca.jsonl
type: alpaca
prompt_style: chatml
- path: datasets/data-oss_instruct-decontaminated-alpaca.jsonl
type: alpaca
prompt_style: chatml
- path: jondurbin/airoboros-3.2
type: sharegpt
conversations: chatml
dataset_prepared_path: last_run_prepared
val_set_size: 0.01
output_dir: ./qlora-out-2
seed: 420
adapter: qlora
lora_model_dir:
sequence_len: 8192
sample_packing: true
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:
lora_modules_to_save:
- embed_tokens
- lm_head
wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:
gradient_accumulation_steps: 2
micro_batch_size: 3
num_epochs: 1.5
optimizer: paged_adamw_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
loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3
warmup_steps: 0
evals_per_epoch: 20
eval_table_size:
saves_per_epoch: 20
debug:
deepspeed:
weight_decay: 0.05
fsdp:
fsdp_config:
special_tokens:
bos_token: "<s>"
eos_token: "</s>"
unk_token: "<unk>"
eos_token: "<|im_end|>"
tokens:
- "<|im_start|>"
trust_remote_code: true
qlora-out-2
This model is a fine-tuned version of rishiraj/CatPPT-base on multiple datasets. It achieves the following results on the evaluation set:
- Loss: 0.4258
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0002
- train_batch_size: 3
- eval_batch_size: 3
- seed: 420
- distributed_type: multi-GPU
- num_devices: 3
- gradient_accumulation_steps: 2
- total_train_batch_size: 18
- total_eval_batch_size: 9
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- num_epochs: 1.5
Training results
| Training Loss | Epoch | Step | Validation Loss |
|---|---|---|---|
| 0.7691 | 0.0 | 1 | 0.7027 |
| 0.5556 | 0.05 | 135 | 0.5808 |
| 0.597 | 0.1 | 270 | 0.5488 |
| 0.5979 | 0.15 | 405 | 0.5277 |
| 0.4693 | 0.2 | 540 | 0.5108 |
| 0.5123 | 0.25 | 675 | 0.4978 |
| 0.4394 | 0.3 | 810 | 0.4883 |
| 0.46 | 0.35 | 945 | 0.4802 |
| 0.4472 | 0.4 | 1080 | 0.4748 |
| 0.47 | 0.45 | 1215 | 0.4687 |
| 0.4249 | 0.5 | 1350 | 0.4637 |
| 0.4823 | 0.55 | 1485 | 0.4599 |
| 0.4209 | 0.6 | 1620 | 0.4555 |
| 0.4909 | 0.65 | 1755 | 0.4517 |
| 0.4663 | 0.7 | 1890 | 0.4470 |
| 0.4215 | 0.75 | 2025 | 0.4437 |
| 0.4267 | 0.8 | 2160 | 0.4398 |
| 0.4109 | 0.85 | 2295 | 0.4364 |
| 0.4099 | 0.9 | 2430 | 0.4331 |
| 0.447 | 0.95 | 2565 | 0.4298 |
| 0.4412 | 1.0 | 2700 | 0.4272 |
| 0.3838 | 1.03 | 2835 | 0.4287 |
| 0.4262 | 1.08 | 2970 | 0.4274 |
| 0.3889 | 1.13 | 3105 | 0.4263 |
| 0.3114 | 1.18 | 3240 | 0.4255 |
| 0.3685 | 1.23 | 3375 | 0.4256 |
| 0.392 | 1.28 | 3510 | 0.4253 |
| 0.3751 | 1.33 | 3645 | 0.4255 |
| 0.3756 | 1.39 | 3780 | 0.4256 |
| 0.3108 | 1.44 | 3915 | 0.4258 |
Framework versions
- Transformers 4.37.0.dev0
- Pytorch 2.1.1+cu121
- Datasets 2.16.1
- Tokenizers 0.15.0
Training procedure
The following bitsandbytes quantization config was used during training:
- quant_method: bitsandbytes
- load_in_8bit: False
- load_in_4bit: True
- 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: nf4
- bnb_4bit_use_double_quant: True
- bnb_4bit_compute_dtype: bfloat16
Framework versions
- PEFT 0.6.0