See axolotl config

axolotl version: 0.4.1

base_model: deepseek-ai/deepseek-coder-1.3b-instruct
# base_model: Qwen/CodeQwen1.5-7B-Chat
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
is_mistral_derived_model: false

load_in_8bit: false
load_in_4bit: true
strict: false

lora_fan_in_fan_out: false
data_seed: 49
seed: 49

datasets:
  - path: sample_data/alpaca_synth_cypher.jsonl
    type: sharegpt
    conversation: alpaca
dataset_prepared_path: last_run_prepared
val_set_size: 0.1
output_dir: ./qlora-alpaca-deepseek-1.3b-inst
# output_dir: ./qlora-alpaca-out

# hub_model_id: jermyn/CodeQwen1.5-7B-Chat-NLQ2Cypher
hub_model_id: jermyn/deepseek-code-1.3b-inst-NLQ2Cypher

adapter: qlora
lora_model_dir:

sequence_len: 896
sample_packing: false
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
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

# If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.
# For LLaMA and Mistral, you need to save `embed_tokens` and `lm_head`. It may vary for other models.
# `embed_tokens` converts tokens to embeddings, and `lm_head` converts embeddings to token probabilities.
# https://github.com/huggingface/peft/issues/334#issuecomment-1561727994
# lora_modules_to_save:
#   - embed_tokens
#   - lm_head

wandb_project: fine-tune-axolotl
wandb_entity: jermyn

gradient_accumulation_steps: 1
micro_batch_size: 16
eval_batch_size: 16
num_epochs: 6
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.001
max_grad_norm: 1.0
adam_beta2: 0.95
adam_epsilon: 0.00001

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: 10
evals_per_epoch: 4
eval_table_size:
eval_table_max_new_tokens: 128
# saves_per_epoch: 6
save_steps: 10
save_total_limit: 3
debug:
weight_decay: 0.0
fsdp:
fsdp_config:
# special_tokens:
#   bos_token: "<s>"
#   eos_token: "</s>"
#   unk_token: "<unk>"
save_safetensors: true

deepseek-code-1.3b-inst-NLQ2Cypher

This model is a fine-tuned version of deepseek-ai/deepseek-coder-1.3b-instruct on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.3816

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: 0.001
• train_batch_size: 16
• eval_batch_size: 16
• seed: 49
• optimizer: Adam with betas=(0.9,0.95) and epsilon=1e-05
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 6

Training results

Training Loss	Epoch	Step	Validation Loss
1.8723	0.1429	1	1.6354
1.9222	0.2857	2	1.5979
1.5222	0.5714	4	1.0739
0.7251	0.8571	6	0.7059
0.3581	1.1429	8	0.6420
0.2804	1.4286	10	0.6576
0.2337	1.7143	12	0.4628
0.1988	2.0	14	0.3674
0.1218	2.2857	16	0.3528
0.1087	2.5714	18	0.3584
0.0842	2.8571	20	0.3569
0.0733	3.1429	22	0.3610
0.0943	3.4286	24	0.3477
0.058	3.7143	26	0.3575
0.0753	4.0	28	0.3797
0.0435	4.2857	30	0.3915
0.0758	4.5714	32	0.3902
0.0608	4.8571	34	0.3875
0.0368	5.1429	36	0.3853
0.0444	5.4286	38	0.3818
0.055	5.7143	40	0.3820
0.0532	6.0	42	0.3816

Framework versions

• PEFT 0.11.1
• Transformers 4.41.1
• Pytorch 2.1.2+cu118
• Datasets 2.19.1
• Tokenizers 0.19.1