base_model: mistralai/Mistral-7B-Instruct-v0.3
# optionally might have model_type or tokenizer_type
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
# Automatically upload checkpoint and final model to HF
hub_model_id: AiAF/Pretrained-QLoRA-r9kilo-V1

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: json
    data_files: ["pretraining.jsonl"]
    type: completion

dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: ./outputs/qlora-out
save_total_limit: 1000

adapter: qlora
lora_model_dir:
lora_r: 256
lora_alpha: 512
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

sequence_len: 512
sample_packing: true
pad_to_sequence_len: true
eval_sample_packing: false

wandb_project: "LLM-Pretraining"
wandb_watch: "all"
wandb_name: "QLoRA-9000-LLM_Datasets-V1"
wandb_log_model: "false"
wandb_run_id: "QLoRA-9000-LLM_Datasets-V1"

gradient_accumulation_steps: 4
micro_batch_size: 64
num_epochs: 4
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.00000047 #0.0000033 #0.000005

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

warmup_steps: 15
evals_per_epoch: 10
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 5
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:

Pretrained-QLoRA-r9kilo-V1

This model is a fine-tuned version of mistralai/Mistral-7B-Instruct-v0.3 on the json dataset. It achieves the following results on the evaluation set:

Loss: 2.0777

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: 4.7e-07
train_batch_size: 64
eval_batch_size: 64
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 256
optimizer: Use OptimizerNames.ADAMW_BNB with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 15
num_epochs: 4.0

Training results

Training Loss	Epoch	Step	Validation Loss
2.4054	0.0126	1	2.3746
2.3966	0.1006	8	2.3649
2.3607	0.2013	16	2.3105
2.2822	0.3019	24	2.2564
2.2658	0.4025	32	2.2115
2.2261	0.5031	40	2.1804
2.1511	0.6038	48	2.1578
2.1752	0.7044	56	2.1431
2.1718	0.8050	64	2.1331
2.1669	0.9057	72	2.1237
2.1408	1.0	80	2.1135
2.1057	1.1006	88	2.1085
2.1289	1.2013	96	2.1038
2.0875	1.3019	104	2.0994
2.1468	1.4025	112	2.0960
2.1295	1.5031	120	2.0933
2.1162	1.6038	128	2.0910
2.1073	1.7044	136	2.0891
2.1002	1.8050	144	2.0875
2.1017	1.9057	152	2.0860
2.0871	2.0	160	2.0849
2.0889	2.1006	168	2.0838
2.1011	2.2013	176	2.0828
2.1061	2.3019	184	2.0820
2.1024	2.4025	192	2.0812
2.1313	2.5031	200	2.0807
2.1128	2.6038	208	2.0801
2.0528	2.7044	216	2.0796
2.1116	2.8050	224	2.0792
2.1395	2.9057	232	2.0789
2.1217	3.0	240	2.0786
2.1046	3.1006	248	2.0784
2.1093	3.2013	256	2.0781
2.1218	3.3019	264	2.0780
2.1058	3.4025	272	2.0779
2.1027	3.5031	280	2.0778
2.0832	3.6038	288	2.0778
2.1026	3.7044	296	2.0777
2.1173	3.8050	304	2.0777
2.1293	3.9057	312	2.0777

Framework versions

PEFT 0.14.0
Transformers 4.49.0
Pytorch 2.5.1+cu124
Datasets 3.2.0
Tokenizers 0.21.0