mnoukhov/pythia410m-dpo2-tldr

README.md

@@ -16,15 +16,15 @@ should probably proofread and complete it, then remove this comment. -->
 
 This model is a fine-tuned version of [mnoukhov/pythia410m-sft-tldr](https://huggingface.co/mnoukhov/pythia410m-sft-tldr) on an unknown dataset.
 It achieves the following results on the evaluation set:
-- Loss: 0.6550
-- Rewards/chosen: -0.1077
-- Rewards/rejected: -0.2150
-- Rewards/accuracies: 0.6269
-- Rewards/margins: 0.1073
-- Logps/rejected: -72.4202
-- Logps/chosen: -72.4202
-- Logps/ref Rejected: -63.5119
-- Logps/ref Chosen: -70.2656
+- Loss: 0.6237
+- Rewards/chosen: -2.4342
+- Rewards/rejected: -2.7481
+- Rewards/accuracies: 0.6489
+- Rewards/margins: 0.3139
+- Logps/rejected: -114.3425
+- Logps/chosen: -114.3425
+- Logps/ref Rejected: -59.5615
+- Logps/ref Chosen: -65.6594
 
 ## Model description
 
@@ -43,25 +43,28 @@ More information needed
 ### Training hyperparameters
 
 The following hyperparameters were used during training:
-- learning_rate: 3e-05
-- train_batch_size: 16
+- learning_rate: 1e-05
+- train_batch_size: 4
 - eval_batch_size: 8
 - seed: 42
 - distributed_type: multi-GPU
+- num_devices: 4
 - gradient_accumulation_steps: 4
 - total_train_batch_size: 64
+- total_eval_batch_size: 32
 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
 - lr_scheduler_type: cosine
 - num_epochs: 1.0
+- mixed_precision_training: Native AMP
 
 ### Training results
 
 | Training Loss | Epoch  | Step | Validation Loss | Rewards/chosen | Rewards/rejected | Rewards/accuracies | Rewards/margins | Logps/rejected | Logps/chosen | Logps/ref Rejected | Logps/ref Chosen |
 |:-------------:|:------:|:----:|:---------------:|:--------------:|:----------------:|:------------------:|:---------------:|:--------------:|:------------:|:------------------:|:----------------:|
-| No log        | 0.2016 | 63   | 0.7000          | -0.0093        | -0.0090          | 0.5043             | -0.0003         | -70.4516       | -70.4516     | -63.5119           | -70.2656         |
-| 0.6682        | 0.4032 | 126  | 0.6714          | -0.0226        | -0.0861          | 0.5957             | 0.0635          | -70.7182       | -70.7182     | -63.5119           | -70.2656         |
-| 0.6682        | 0.6048 | 189  | 0.6623          | -0.0820        | -0.1707          | 0.6142             | 0.0887          | -71.9055       | -71.9055     | -63.5119           | -70.2656         |
-| 0.6423        | 0.8064 | 252  | 0.6550          | -0.1077        | -0.2150          | 0.6269             | 0.1073          | -72.4202       | -72.4202     | -63.5119           | -70.2656         |
+| No log        | 0.2016 | 63   | 0.6296          | -0.3668        | -0.5301          | 0.6780             | 0.1633          | -72.9961       | -72.9961     | -59.5615           | -65.6594         |
+| 0.5884        | 0.4032 | 126  | 0.6180          | -1.5625        | -1.8194          | 0.6596             | 0.2569          | -96.9089       | -96.9089     | -59.5615           | -65.6594         |
+| 0.5884        | 0.6048 | 189  | 0.6302          | -2.3096        | -2.5816          | 0.6445             | 0.2720          | -111.8520      | -111.8520    | -59.5615           | -65.6594         |
+| 0.4615        | 0.8064 | 252  | 0.6237          | -2.4342        | -2.7481          | 0.6489             | 0.3139          | -114.3425      | -114.3425    | -59.5615           | -65.6594         |
 
 
 ### Framework versions

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ad6230f86f228c40c1e6d7bc9fe8fb2d0bbdd0accb1065d02ed94e674d97f81f
+oid sha256:69b6220103f85750164fc3f8c7dc64d08043a12b521325782ae63aef087193e6
 size 25192592