update config embeddings

config.json

@@ -11,7 +11,7 @@
   "initializer_range": 0.02,
   "intermediate_size": 16384,
   "layer_norm_eps": 1e-05,
-  "max_position_embeddings": 2048,
+  "max_position_embeddings": 8192,
   "model_type": "gpt_neox",
   "num_attention_heads": 32,
   "num_hidden_layers": 32,

configs/minotaur.yml

@@ -0,0 +1,107 @@
+base_model: /workspace/pythia-6.9b-deduped-8k/
+base_model_config: EleutherAI/pythia-6.9b-deduped
+model_type: GPTNeoXForCausalLM
+tokenizer_type: AutoTokenizer
+load_in_8bit: false
+load_in_4bit: false
+strict: False
+gptq: false
+device_map: auto
+push_dataset_to_hub: winglian
+hf_use_auth_token: true
+  # dataset_shard_num: 3
+  # dataset_shard_idx: 0
+datasets:
+  - path: winglian/pygmalion-cleaned
+    data_files:
+      - v13_no_ai.cleaned.jsonl
+    type: pygmalion
+  - path: winglian/evals
+    data_files:
+      - hf/ARC-Challenge.jsonl
+      - hf/ARC-Easy.jsonl
+      - hf/riddle_sense.jsonl
+    type: explainchoice:chat
+  - path: winglian/evals
+    data_files:
+      - openai/tldr.jsonl
+    type: summarizetldr:chat
+  - path: winglian/evals
+    data_files:
+      - hf/gsm8k.jsonl
+    type: alpaca_chat.load_qa
+  - path: winglian/evals
+    data_files:
+      - hellaswag/hellaswag.jsonl
+    type: explainchoice:chat
+  - path: metaeval/ScienceQA_text_only
+    type: concisechoice:chat
+  - path: ehartford/WizardLM_alpaca_evol_instruct_70k_unfiltered
+    type: alpaca_chat
+  - path: ehartford/wizard_vicuna_70k_unfiltered
+    type: sharegpt:chat
+  - path: winglian/chatlogs-en-cleaned
+    data_files:
+      - sharegpt_cleaned.jsonl
+    type: sharegpt:chat
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca_chat
+  - path: teknium/GPTeacher-General-Instruct
+    data_files: gpt4-instruct-similarity-0.6-dataset.json
+    type: gpteacher:chat
+  - path: ewof/code-alpaca-instruct-unfiltered
+    type: alpaca_chat
+  - 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_chat
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+adapter:
+lora_model_dir:
+sequence_len: 8192
+max_packed_sequence_len: 8192
+lora_r: 64
+lora_alpha: 32
+lora_dropout: 0.0
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out: true  # pythia/GPTNeoX lora specific
+wandb_project: minotaur-7b
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+output_dir: ./minotaur-7b
+gradient_accumulation_steps: 1
+micro_batch_size: 4
+num_epochs: 1
+learning_rate: 0.00003
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+warmup_steps: 100
+save_steps: 84
+eval_steps: 10
+train_on_inputs: false
+group_by_length: true
+bf16: false
+fp16: false
+bfloat16: true
+tf32: true
+flash_optimum: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+gradient_checkpointing: true
+fsdp:
+fsdp_transformer_layer_cls_to_wrap:
+collator_pad_to_longest: true
+save_steps: 35