Create TM-Mistral-Config.yaml

TM-Mistral-Config.yaml

@@ -0,0 +1,84 @@
+
+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/QLoRA-Finetune-LMTV-TM-9-2320-365-10
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: AiAF/TM_plain_qa_list_with_special_tokens.jsonl
+    ds_type: json
+    type: chat_template
+    chat_template: chatml
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+    roles:
+      user:
+        - human
+      assistant:
+        - gpt
+      system:
+        - system
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./TM-9-2320-365-10_V10
+
+save_total_limit: 100
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+pad_to_sequence_len: true
+
+adapter: qlora
+lora_r: 256
+lora_alpha: 512
+lora_dropout: 0.05
+#lora_target_linear: true
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+
+wandb_project: "LLM-Pretraining"
+wandb_watch: "all"
+wandb_name: "LMTV-TM-V10"
+wandb_log_model: "false"
+wandb_run_id: "LMTV-TM-V10"
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 10
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.000005
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
+warmup_steps: 10
+evals_per_epoch: 5
+saves_per_epoch: 5
+weight_decay: 0.0
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"