feat: train with sorted label to make training behavior reproducible at evaluation time

README.md

@@ -16,7 +16,7 @@ should probably proofread and complete it, then remove this comment. -->
 
 This model is a fine-tuned version of [openai-gpt](https://huggingface.co/openai-gpt) on an unknown dataset.
 It achieves the following results on the evaluation set:
-- Loss: 0.3986
+- Loss: 0.3295
 
 ## Model description
 
@@ -47,13 +47,14 @@ The following hyperparameters were used during training:
 
 | Training Loss | Epoch | Step | Validation Loss |
 |:-------------:|:-----:|:----:|:---------------:|
-| 1.5016        | 1.0   | 12   | 1.1160          |
-| 0.8736        | 2.0   | 24   | 0.7228          |
-| 0.4681        | 3.0   | 36   | 0.4849          |
-| 0.2851        | 4.0   | 48   | 0.4339          |
-| 0.1733        | 5.0   | 60   | 0.3113          |
-| 0.1108        | 6.0   | 72   | 0.3303          |
-| 0.0841        | 7.0   | 84   | 0.3986          |
+| 1.6661        | 1.0   | 12   | 1.4575          |
+| 1.3696        | 2.0   | 24   | 1.0023          |
+| 0.6977        | 3.0   | 36   | 0.5564          |
+| 0.3751        | 4.0   | 48   | 0.4573          |
+| 0.2528        | 5.0   | 60   | 0.3803          |
+| 0.1503        | 6.0   | 72   | 0.2699          |
+| 0.0863        | 7.0   | 84   | 0.2888          |
+| 0.0547        | 8.0   | 96   | 0.3295          |
 
 
 ### Framework versions

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7c3608554d7067516442941a29037b9956dca9679dce7edb8295503bab207489
+oid sha256:5e13a56823f51b105f62aa5a86397d4e9c9a21aa276bd6e12918cfcf65b2a0e7
 size 466181672