feat: training run with increased lora rank, application, and lora alpha

README.md

@@ -16,7 +16,7 @@ should probably proofread and complete it, then remove this comment. -->
 
 This model is a fine-tuned version of [google/gemma-3-1b-pt](https://huggingface.co/google/gemma-3-1b-pt) on an unknown dataset.
 It achieves the following results on the evaluation set:
-- Loss: 0.4730
+- Loss: 0.5270
 
 ## Model description
 
@@ -35,35 +35,30 @@ More information needed
 ### Training hyperparameters
 
 The following hyperparameters were used during training:
-- learning_rate: 0.0001
-- train_batch_size: 8
-- eval_batch_size: 8
+- learning_rate: 6.5e-05
+- train_batch_size: 16
+- eval_batch_size: 16
 - seed: 42
 - gradient_accumulation_steps: 4
-- total_train_batch_size: 32
+- total_train_batch_size: 64
 - optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
 - lr_scheduler_type: linear
-- num_epochs: 15
+- num_epochs: 20
 
 ### Training results
 
 | Training Loss | Epoch | Step | Validation Loss |
 |:-------------:|:-----:|:----:|:---------------:|
-| 2.1658        | 1.0   | 24   | 2.0002          |
-| 1.8154        | 2.0   | 48   | 1.5934          |
-| 1.3517        | 3.0   | 72   | 1.0746          |
-| 0.8524        | 4.0   | 96   | 0.6859          |
-| 0.6272        | 5.0   | 120  | 0.5927          |
-| 0.5657        | 6.0   | 144  | 0.5468          |
-| 0.5226        | 7.0   | 168  | 0.5092          |
-| 0.4984        | 8.0   | 192  | 0.4978          |
-| 0.489         | 9.0   | 216  | 0.4904          |
-| 0.482         | 10.0  | 240  | 0.4851          |
-| 0.4766        | 11.0  | 264  | 0.4806          |
-| 0.4722        | 12.0  | 288  | 0.4773          |
-| 0.4693        | 13.0  | 312  | 0.4751          |
-| 0.467         | 14.0  | 336  | 0.4736          |
-| 0.4657        | 15.0  | 360  | 0.4730          |
+| 1.6969        | 1.0   | 12   | 0.9517          |
+| 0.7666        | 2.0   | 24   | 0.6638          |
+| 0.5743        | 3.0   | 36   | 0.5845          |
+| 0.4948        | 4.0   | 48   | 0.5431          |
+| 0.436         | 5.0   | 60   | 0.5176          |
+| 0.4011        | 6.0   | 72   | 0.5075          |
+| 0.3743        | 7.0   | 84   | 0.5028          |
+| 0.3507        | 8.0   | 96   | 0.5099          |
+| 0.3313        | 9.0   | 108  | 0.5267          |
+| 0.3121        | 10.0  | 120  | 0.5270          |
 
 
 ### Framework versions

adapter_config.json

@@ -13,18 +13,23 @@
   "layers_pattern": null,
   "layers_to_transform": null,
   "loftq_config": {},
-  "lora_alpha": 8,
+  "lora_alpha": 128,
   "lora_bias": false,
-  "lora_dropout": 0.0,
+  "lora_dropout": 0.05,
   "megatron_config": null,
   "megatron_core": "megatron.core",
   "modules_to_save": null,
   "peft_type": "LORA",
-  "r": 8,
+  "r": 64,
   "rank_pattern": {},
   "revision": null,
   "target_modules": [
+    "up_proj",
+    "down_proj",
+    "gate_proj",
+    "k_proj",
     "q_proj",
+    "o_proj",
     "v_proj"
   ],
   "task_type": "CAUSAL_LM",

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:87d8f68c8f6c2d9f20c8bf09e9f637b4f1fa51610c91bac87203015f3e99e5b0
-size 2995512
+oid sha256:ca16fca62c89606b182d92af1ed21f074b5d55c904bd235fc61590f1c4639f0e
+size 208780928

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:85d04b7842181c0e4f6d2b93c8b56af00eb900b9352bab086d1a999b7e61ed88
+oid sha256:72b5334e34efcb5c7b466960624b42c17554302ae914dac04c033c2b588ba158
 size 5368