End of training

README.md

@@ -41,12 +41,12 @@ early_stopping_patience: null
 eval_max_new_tokens: 128
 eval_table_size: null
 evals_per_epoch: 4
-flash_attention: true
-fp16: true
+flash_attention: false
+fp16: null
 fsdp: null
 fsdp_config: null
 gradient_accumulation_steps: 4
-gradient_checkpointing: true
+gradient_checkpointing: false
 group_by_length: false
 hub_model_id: dimasik87/8102bae9-c442-4a29-848f-946f8771f1f9
 hub_repo: null
@@ -57,28 +57,28 @@ load_in_4bit: false
 load_in_8bit: false
 local_rank: null
 logging_steps: 1
-lora_alpha: 16
-lora_dropout: 0.1
+lora_alpha: 32
+lora_dropout: 0.05
 lora_fan_in_fan_out: null
 lora_model_dir: null
-lora_r: 8
+lora_r: 16
 lora_target_linear: true
 lr_scheduler: cosine
 max_memory:
   0: 70GiB
-max_steps: 25
-micro_batch_size: 1
+max_steps: 50
+micro_batch_size: 2
 mlflow_experiment_name: /tmp/5daf839d73ce7025_train_data.json
 model_type: AutoModelForCausalLM
-num_epochs: 3
-optimizer: adamw_torch
+num_epochs: 4
+optimizer: adamw_bnb_8bit
 output_dir: miner_id_24
 pad_to_sequence_len: true
 resume_from_checkpoint: null
 s2_attention: null
 sample_packing: false
-saves_per_epoch: 3
-sequence_len: 2028
+saves_per_epoch: 4
+sequence_len: 1024
 strict: false
 tf32: false
 tokenizer_type: AutoTokenizer
@@ -103,7 +103,7 @@ xformers_attention: null
 
 This model is a fine-tuned version of [unsloth/Hermes-3-Llama-3.1-8B](https://huggingface.co/unsloth/Hermes-3-Llama-3.1-8B) on the None dataset.
 It achieves the following results on the evaluation set:
-- Loss: 0.6604
+- Loss: nan
 
 ## Model description
 
@@ -123,29 +123,33 @@ More information needed
 
 The following hyperparameters were used during training:
 - learning_rate: 0.0002
-- train_batch_size: 1
-- eval_batch_size: 1
+- train_batch_size: 2
+- eval_batch_size: 2
 - seed: 42
 - gradient_accumulation_steps: 4
-- total_train_batch_size: 4
-- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
+- total_train_batch_size: 8
+- optimizer: Use OptimizerNames.ADAMW_BNB with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
 - lr_scheduler_type: cosine
 - lr_scheduler_warmup_steps: 10
-- training_steps: 25
+- training_steps: 50
 
 ### Training results
 
 | Training Loss | Epoch  | Step | Validation Loss |
 |:-------------:|:------:|:----:|:---------------:|
-| 6.2351        | 0.0004 | 1    | 7.1638          |
-| 7.0317        | 0.0013 | 3    | 7.1373          |
-| 5.6858        | 0.0025 | 6    | 6.6630          |
-| 5.0704        | 0.0038 | 9    | 4.9700          |
-| 3.2353        | 0.0051 | 12   | 2.9853          |
-| 1.8729        | 0.0063 | 15   | 1.5725          |
-| 0.8863        | 0.0076 | 18   | 0.9877          |
-| 1.4191        | 0.0089 | 21   | 0.7656          |
-| 0.397         | 0.0101 | 24   | 0.6604          |
+| 0.0           | 0.0008 | 1    | nan             |
+| 0.0           | 0.0034 | 4    | nan             |
+| 0.0           | 0.0068 | 8    | nan             |
+| 0.0           | 0.0101 | 12   | nan             |
+| 0.0           | 0.0135 | 16   | nan             |
+| 0.0           | 0.0169 | 20   | nan             |
+| 0.0           | 0.0203 | 24   | nan             |
+| 0.0           | 0.0236 | 28   | nan             |
+| 0.0           | 0.0270 | 32   | nan             |
+| 0.0           | 0.0304 | 36   | nan             |
+| 0.0           | 0.0338 | 40   | nan             |
+| 0.0           | 0.0371 | 44   | nan             |
+| 0.0           | 0.0405 | 48   | nan             |
 
 
 ### Framework versions

adapter_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ec2f9c10a826b35d8c76e3d52568ff4b5c53dbea3b5c23956e2103885b037f5c
-size 84047370
+oid sha256:45fe9746a7db9cd15165b811a5331f0dc4b1d8ec660979d5129957450fdbfea6
+size 167934026