106 lines
2.8 KiB
YAML
106 lines
2.8 KiB
YAML
base_model: huggyllama/llama-13b
|
|
# base_model: /workspace/manticore-13b/
|
|
base_model_config: huggyllama/llama-13b
|
|
model_type: LlamaForCausalLM
|
|
tokenizer_type: LlamaTokenizer
|
|
load_in_8bit: false
|
|
datasets:
|
|
- path: winglian/evals
|
|
data_files:
|
|
- hf/ARC-Challenge.jsonl
|
|
- hf/ARC-Easy.jsonl
|
|
- mmlu/abstract_algebra.jsonl
|
|
- mmlu/conceptual_physics.jsonl
|
|
- mmlu/formal_logic.jsonl
|
|
- mmlu/high_school_physics.jsonl
|
|
- mmlu/logical_fallacies.jsonl
|
|
type: explainchoice
|
|
- path: winglian/evals
|
|
data_files:
|
|
- openai/tldr.jsonl
|
|
type: summarizetldr
|
|
- path: winglian/evals
|
|
data_files:
|
|
- hellaswag/hellaswag-concise.jsonl
|
|
type: concisechoice
|
|
- path: metaeval/ScienceQA_text_only
|
|
type: concisechoice
|
|
- path: ehartford/WizardLM_alpaca_evol_instruct_70k_unfiltered
|
|
type: alpaca
|
|
- path: ehartford/wizard_vicuna_70k_unfiltered
|
|
type: sharegpt
|
|
- path: winglian/chatlogs-en-cleaned
|
|
data_files:
|
|
- sharegpt_cleaned.jsonl
|
|
type: sharegpt
|
|
- path: teknium/GPT4-LLM-Cleaned
|
|
type: alpaca
|
|
- path: teknium/GPTeacher-General-Instruct
|
|
data_files: gpt4-instruct-similarity-0.6-dataset.json
|
|
type: gpteacher
|
|
- path: QingyiSi/Alpaca-CoT
|
|
data_files:
|
|
- Chain-of-Thought/formatted_cot_data/aqua_train.json
|
|
- Chain-of-Thought/formatted_cot_data/creak_train.json
|
|
- Chain-of-Thought/formatted_cot_data/ecqa_train.json
|
|
- Chain-of-Thought/formatted_cot_data/esnli_train.json
|
|
- Chain-of-Thought/formatted_cot_data/gsm8k_train.json
|
|
- Chain-of-Thought/formatted_cot_data/qasc_train.json
|
|
- Chain-of-Thought/formatted_cot_data/qed_train.json
|
|
- Chain-of-Thought/formatted_cot_data/sensemaking_train.json
|
|
- Chain-of-Thought/formatted_cot_data/strategyqa_train.json
|
|
- GPTeacher/Roleplay/formatted_roleplay-similarity_0.6-instruct-dataset.json
|
|
type: alpaca
|
|
dataset_prepared_path: last_run_prepared
|
|
val_set_size: 0.02
|
|
adapter:
|
|
lora_model_dir:
|
|
sequence_len: 2048
|
|
max_packed_sequence_len: 2048
|
|
lora_r:
|
|
lora_alpha:
|
|
lora_dropout:
|
|
lora_target_modules:
|
|
lora_fan_in_fan_out:
|
|
wandb_project: manticore-13b
|
|
wandb_watch:
|
|
wandb_run_id:
|
|
wandb_log_model:
|
|
output_dir: ./manticore-13b
|
|
batch_size: 512
|
|
micro_batch_size: 8
|
|
num_epochs: 4
|
|
optimizer:
|
|
torchdistx_path:
|
|
lr_scheduler:
|
|
learning_rate: 0.000032
|
|
train_on_inputs: false
|
|
group_by_length: false
|
|
bf16: true
|
|
tf32: true
|
|
gradient_checkpointing: true
|
|
early_stopping_patience:
|
|
resume_from_checkpoint:
|
|
local_rank:
|
|
logging_steps: 1
|
|
xformers_attention: true
|
|
flash_attention:
|
|
gptq_groupsize:
|
|
gptq_model_v1:
|
|
warmup_steps: 20
|
|
eval_steps: 10
|
|
save_steps:
|
|
debug:
|
|
deepspeed:
|
|
weight_decay: 0
|
|
fsdp:
|
|
- full_shard
|
|
- auto_wrap
|
|
fsdp_config:
|
|
fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
|
|
special_tokens:
|
|
bos_token: "<s>"
|
|
eos_token: "</s>"
|
|
unk_token: "<unk>"
|
|
|