AiAF/Pretraining-Yellow-Atol-7B-Instruct-V1

See axolotl config

axolotl version: 0.8.0.dev0

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/Pretraining-SpongeBoB-7B-Instruct-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: 0.1
output_dir: ./outputs/qlora-out/Pretraining-SpongeBoB-7B-Instruct-V1
save_total_limit: 10

adapter: qlora
lora_model_dir:

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

lora_r: 256
lora_alpha: 64
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

wandb_project: "LLM-Pretraining"
wandb_entity:
wandb_watch: "all"
wandb_name: "Pretraining-SpongeBoB-7B-Instruct-V1"
wandb_run_id: "Pretraining-SpongeBoB-7B-Instruct-V1"
wandb_log_model: "false"

gradient_accumulation_steps: 2
micro_batch_size: 9
num_epochs: 10
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 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

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

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

Pretraining-SpongeBoB-7B-Instruct-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: 1.6255

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: 5e-06
• train_batch_size: 9
• eval_batch_size: 9
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 18
• 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: 10
• num_epochs: 10.0

Training results

Training Loss	Epoch	Step	Validation Loss
1.7843	0.0417	1	1.7939
1.8262	0.2083	5	1.7915
1.839	0.4167	10	1.7733
1.7503	0.625	15	1.7438
1.7191	0.8333	20	1.7260
1.7191	1.0417	25	1.7138
1.7548	1.25	30	1.7023
1.6795	1.4583	35	1.6924
1.6848	1.6667	40	1.6836
1.6856	1.875	45	1.6770
1.7155	2.0833	50	1.6715
1.6901	2.2917	55	1.6665
1.6797	2.5	60	1.6621
1.6704	2.7083	65	1.6581
1.6763	2.9167	70	1.6545
1.678	3.125	75	1.6516
1.6271	3.3333	80	1.6490
1.662	3.5417	85	1.6468
1.6384	3.75	90	1.6446
1.6273	3.9583	95	1.6427
1.5934	4.1667	100	1.6408
1.6217	4.375	105	1.6393
1.6383	4.5833	110	1.6378
1.6244	4.7917	115	1.6365
1.6238	5.0	120	1.6352
1.6179	5.2083	125	1.6340
1.6203	5.4167	130	1.6330
1.6177	5.625	135	1.6319
1.6332	5.8333	140	1.6310
1.6277	6.0417	145	1.6302
1.6461	6.25	150	1.6296
1.6668	6.4583	155	1.6290
1.6249	6.6667	160	1.6284
1.6013	6.875	165	1.6278
1.6098	7.0833	170	1.6274
1.5954	7.2917	175	1.6270
1.6488	7.5	180	1.6267
1.6153	7.7083	185	1.6264
1.6232	7.9167	190	1.6262
1.6611	8.125	195	1.6260
1.5997	8.3333	200	1.6258
1.6166	8.5417	205	1.6258
1.6427	8.75	210	1.6256
1.6157	8.9583	215	1.6255
1.6303	9.1667	220	1.6255
1.6179	9.375	225	1.6255
1.6063	9.5833	230	1.6255
1.6043	9.7917	235	1.6255
1.5881	10.0	240	1.6255

Framework versions

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