Update README.md

README.md

@@ -14,4 +14,56 @@ Write a story using this writing prompt: As a prank a witch detached your cock a
 
 Apparently RP has also become a bit less sloppy by coincidence.
 
-We are looking into opening the datasets up, I'm a bit tired atm, you can also just go get this torrent of the entire reddit, select only the subreddits you want and DIY [https://academictorrents.com/details/56aa49f9653ba545f48df2e33679f014d2829c10]
+We are looking into opening the datasets up, I'm a bit tired atm, you can also just go get this torrent of the entire reddit, select only the subreddits you want and DIY [https://academictorrents.com/details/56aa49f9653ba545f48df2e33679f014d2829c10]
+
+(for context - this model was a test run, on a small dataset. It will be scaled up later.) 
+
+
+## Training Config:
+
+Thanks a lot to llamafactory, the easiest train I've ever done so far.
+
+```
+llamafactory-cli train \
+    --stage kto \
+    --do_train True \
+    --model_name_or_path cognitivecomputations/dolphin-2.9.1-llama-3-8b \
+    --preprocessing_num_workers 16 \
+    --finetuning_type lora \
+    --quantization_bit 8 \
+    --template chatml \
+    --flash_attn auto \
+    --use_unsloth True \
+    --dataset_dir /workspace/kto \
+    --dataset kto_dataset \
+    --cutoff_len 2048 \
+    --learning_rate 5e-05 \
+    --num_train_epochs 3.0 \
+    --max_samples 100000 \
+    --per_device_train_batch_size 2 \
+    --gradient_accumulation_steps 8 \
+    --lr_scheduler_type cosine \
+    --max_grad_norm 1.0 \
+    --logging_steps 5 \
+    --save_steps 500 \
+    --warmup_steps 50 \
+    --optim adamw_torch \
+    --packing False \
+    --report_to all \
+    --output_dir saves/LLaMA3-8B/lora/train_2024-06-15-15-18-25 \
+    --bf16 True \
+    --plot_loss True \
+    --ddp_timeout 180000000 \
+    --include_num_input_tokens_seen True \
+    --lora_rank 32 \
+    --lora_alpha 32 \
+    --lora_dropout 0 \
+    --lora_target all \
+    --pref_beta 0.1 \
+    --pref_ftx 0 \
+    --pref_loss sigmoid \
+    --val_size 0.05 \
+    --eval_strategy steps \
+    --eval_steps 50 \
+    --per_device_eval_batch_size 2
+```