0.4.11base_model: EleutherAI/pythia-160m-deduped
2load_in_8bit:
3datasets:
4 - path: jtatman/storywriting_combined_instruct
5 type: alpaca
6dataset_prepared_path: ds-storytelling
7chat_template: inst
8val_set_size: 0.01
9adapter: lora
10lora_model_dir:
11sequence_len: 2048
12lora_r: 16
13lora_alpha: 32
14lora_dropout: 0.05
15lora_target_modules:
16 - query_key_value
17lora_target_linear: true
18lora_fan_in_fan_out: true # pythia/GPTNeoX lora specific
19lora_modules_to_save:
20 - embed_in
21 - embed_out
22 - lm_head
23lora_on_cpu: false
24# ReLoRA configuration
25# # Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed
26# relora_steps: # Number of steps per ReLoRA restart
27# relora_warmup_steps: # Number of per-restart warmup steps
28# relora_anneal_steps: # Number of anneal steps for each relora cycle
29# relora_prune_ratio: # threshold for optimizer magnitude when pruning
30# relora_cpu_offload: # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings
31relora_steps: 200
32relora_warmup_steps: 10
33relora_cpu_offload: false
34wandb_project: pythia
35wandb_entity:
36wandb_watch:
37wandb_name: pythia-160m-storytelling
38wandb_log_model:
39output_dir: ./outputs/lora-alpaca-pythia-160m-storytelling
40gradient_accumulation_steps: 16
41micro_batch_size: 1
42num_epochs: 3
43learning_rate: 0.004
44lr_scheduler: cosine_with_restarts
45#cosine_min_lr_ratio: 0.1
46train_on_inputs: false
47group_by_length: false
48#bf16: auto
49#fp16: true
50#tf32: false
51float16: true
52flash_attn:
53xformers_attention: true
54optimizer: paged_adamw_8bit
55gpu_memory_limit: 8GiB
56hub_model_id: jtatman/pythia-160m-storytelling
57early_stopping_patience: 3
58#resume_from_checkpoint: outputs/lora-alpaca-pythia-125m/checkpoint-51040
59auto_resume_from_checkpoints: true
60local_rank:
61weight_decay: 0.0
62#evals_per_epoch: 4
63eval_steps: 200
64logging_steps: 1
65save_steps: 200
66save_total_limit: 5
67warmup_steps: 100
68tokens:
69 - "[INST]"
70 - "[/INST]"
71| Training Loss | Epoch | Step | Validation Loss |
|---|---|---|---|
| 5.5185 | 0.0012 | 1 | 4.8238 |
| 4.2012 | 0.2348 | 200 | 4.1556 |
| 4.4185 | 0.4696 | 400 | 4.8159 |
| 5.0973 | 0.7043 | 600 | 5.0363 |
| 8.1159 | 0.9391 | 800 | 8.4966 |
| 6.7656 | 1.1739 | 1000 | 7.1575 |
| 7.0548 | 1.4087 | 1200 | 7.3539 |
| 5.9982 | 1.6445 | 1400 | 5.9954 |
| 5.7662 | 1.8792 | 1600 | 6.0222 |
| 4.8094 | 2.1140 | 1800 | 5.0097 |
"Open LLM Leaderboard": {
"exact_match,flexible-extract": 0.022,
"exact_match_stderr,flexible-extract": 0.006566447781940106,
"acc_norm,none": 0.318,
"acc_norm_stderr,none": 0.014487919091408506,
"acc,none": 0.2664044125478186,
"acc_stderr,none": 0.003623534644130716,
"bleu_diff,none": -0.6500479549286462,
"bleu_diff_stderr,none": 0.6420841882903697,
"rougeL_diff,none": -0.7765084899781842,
"rougeL_diff_stderr,none": 1.0033586571635116,
"exact_match,strict-match": 0.006,
"exact_match_stderr,strict-match": 0.003457152557758373,
"rouge2_acc,none": 0.192,
"rouge2_acc_stderr,none": 0.017632180454360994,
"rouge1_acc,none": 0.37,
"rouge1_acc_stderr,none": 0.02161328916516578,
"bleu_acc,none": 0.436,
"bleu_acc_stderr,none": 0.0221989546414768,
"rouge1_diff,none": -1.5563905118333812,
"rouge1_diff_stderr,none": 1.022327995054994,
"rouge2_diff,none": -3.3177627227020277,
"rouge2_diff_stderr,none": 0.9477297777821475,
"bleu_max,none": 15.229235419512532,
"bleu_max_stderr,none": 0.6713582602539528,
"rouge2_max,none": 16.487324929036955,
"rouge2_max_stderr,none": 1.0171593586088354,
"rouge1_max,none": 36.3549677399668,
"rouge1_max_stderr,none": 0.9461627463383844,
"rougeL_max,none": 33.87976960164143,
"rougeL_max_stderr,none": 0.9366539036852334,
"rougeL_acc,none": 0.386,
"rougeL_acc_stderr,none": 0.021793529219281158,
"alias": "Open LLM Leaderboard"
},