wip deepseek train

training/config.yaml

@@ -4,6 +4,11 @@ model:
 params:
   learning_rate: 1e-5
   epochs: 3
+  max_seq_length: 1024
+  quantization: "4bit"  # Options: "none", "4bit", "8bit"
+  per_device_batch_size: 1
+  gradient_accumulation_steps: 8
+  mixed_precision: "no"  # Options: "no", "fp16", "bf16"
 data:
   - "/root/sia/training/clean_start/"
   - "/root/sia/training/delete_indicated_entries/"