Model save

README.md

@@ -1,55 +1,69 @@
 ---
-library_name: peft
-license: llama3.2
 base_model: meta-llama/Llama-3.2-3B-Instruct
+library_name: transformers
+model_name: Llama-3.2-3B-DPO
 tags:
+- generated_from_trainer
 - trl
 - dpo
-- generated_from_trainer
-model-index:
-- name: Llama-3.2-3B-DPO
-  results: []
+licence: license
 ---
 
-<!-- This model card has been generated automatically according to the information the Trainer had access to. You
-should probably proofread and complete it, then remove this comment. -->
+# Model Card for Llama-3.2-3B-DPO
 
-[<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="200" height="32"/>](https://wandb.ai/sciarrilli/dpo-llama32/runs/degfzo4x)
-# Llama-3.2-3B-DPO
+This model is a fine-tuned version of [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct).
+It has been trained using [TRL](https://github.com/huggingface/trl).
 
-This model is a fine-tuned version of [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) on an unknown dataset.
+## Quick start
 
-## Model description
+```python
+from transformers import pipeline
 
-More information needed
+question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
+generator = pipeline("text-generation", model="sciarrilli/Llama-3.2-3B-DPO", device="cuda")
+output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
+print(output["generated_text"])
+```
 
-## Intended uses & limitations
+## Training procedure
 
-More information needed
+[<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="150" height="24"/>](https://wandb.ai/sciarrilli/dpo-llama32/runs/degfzo4x) 
 
-## Training and evaluation data
 
-More information needed
+This model was trained with DPO, a method introduced in [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://huggingface.co/papers/2305.18290).
 
-## Training procedure
+### Framework versions
 
-### Training hyperparameters
+- TRL: 0.15.2
+- Transformers: 4.49.0
+- Pytorch: 2.6.0+cu126
+- Datasets: 3.4.1
+- Tokenizers: 0.21.1
 
-The following hyperparameters were used during training:
-- learning_rate: 5e-06
-- train_batch_size: 2
-- eval_batch_size: 8
-- seed: 42
-- gradient_accumulation_steps: 8
-- total_train_batch_size: 16
-- 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: 1.0
+## Citations
 
-### Framework versions
+Cite DPO as:
+
+```bibtex
+@inproceedings{rafailov2023direct,
+    title        = {{Direct Preference Optimization: Your Language Model is Secretly a Reward Model}},
+    author       = {Rafael Rafailov and Archit Sharma and Eric Mitchell and Christopher D. Manning and Stefano Ermon and Chelsea Finn},
+    year         = 2023,
+    booktitle    = {Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023},
+    url          = {http://papers.nips.cc/paper_files/paper/2023/hash/a85b405ed65c6477a4fe8302b5e06ce7-Abstract-Conference.html},
+    editor       = {Alice Oh and Tristan Naumann and Amir Globerson and Kate Saenko and Moritz Hardt and Sergey Levine},
+}
+```
 
-- PEFT 0.14.0
-- Transformers 4.49.0
-- Pytorch 2.6.0+cu126
-- Datasets 3.4.1
-- Tokenizers 0.21.1
+Cite TRL as:
+    
+```bibtex
+@misc{vonwerra2022trl,
+	title        = {{TRL: Transformer Reinforcement Learning}},
+	author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallouédec},
+	year         = 2020,
+	journal      = {GitHub repository},
+	publisher    = {GitHub},
+	howpublished = {\url{https://github.com/huggingface/trl}}
+}
+```

all_results.json

@@ -1,15 +1,15 @@
 {
-    "epoch": 0.9990344383649823,
-    "eval_logits/chosen": 0.6598725914955139,
-    "eval_logits/rejected": 0.470087468624115,
-    "eval_logps/chosen": -349.9416198730469,
-    "eval_logps/rejected": -318.07635498046875,
-    "eval_loss": 0.6646144986152649,
-    "eval_rewards/accuracies": 0.6153846383094788,
-    "eval_rewards/chosen": -0.0643770769238472,
-    "eval_rewards/margins": 0.06126875430345535,
-    "eval_rewards/rejected": -0.12564583122730255,
-    "eval_runtime": 85.284,
-    "eval_samples_per_second": 1.173,
-    "eval_steps_per_second": 0.152
+    "epoch": 0.9957081545064378,
+    "eval_logits/chosen": 0.6785366535186768,
+    "eval_logits/rejected": 0.5101125240325928,
+    "eval_logps/chosen": -353.8060607910156,
+    "eval_logps/rejected": -244.0733184814453,
+    "eval_loss": 0.6846491098403931,
+    "eval_rewards/accuracies": 0.625,
+    "eval_rewards/chosen": -0.03933782875537872,
+    "eval_rewards/margins": 0.03162214532494545,
+    "eval_rewards/rejected": -0.07095997035503387,
+    "eval_runtime": 9.0831,
+    "eval_samples_per_second": 1.101,
+    "eval_steps_per_second": 0.22
 }

eval_results.json

@@ -1,15 +1,15 @@
 {
-    "epoch": 0.9990344383649823,
-    "eval_logits/chosen": 0.6598725914955139,
-    "eval_logits/rejected": 0.470087468624115,
-    "eval_logps/chosen": -349.9416198730469,
-    "eval_logps/rejected": -318.07635498046875,
-    "eval_loss": 0.6646144986152649,
-    "eval_rewards/accuracies": 0.6153846383094788,
-    "eval_rewards/chosen": -0.0643770769238472,
-    "eval_rewards/margins": 0.06126875430345535,
-    "eval_rewards/rejected": -0.12564583122730255,
-    "eval_runtime": 85.284,
-    "eval_samples_per_second": 1.173,
-    "eval_steps_per_second": 0.152
+    "epoch": 0.9957081545064378,
+    "eval_logits/chosen": 0.6785366535186768,
+    "eval_logits/rejected": 0.5101125240325928,
+    "eval_logps/chosen": -353.8060607910156,
+    "eval_logps/rejected": -244.0733184814453,
+    "eval_loss": 0.6846491098403931,
+    "eval_rewards/accuracies": 0.625,
+    "eval_rewards/chosen": -0.03933782875537872,
+    "eval_rewards/margins": 0.03162214532494545,
+    "eval_rewards/rejected": -0.07095997035503387,
+    "eval_runtime": 9.0831,
+    "eval_samples_per_second": 1.101,
+    "eval_steps_per_second": 0.22
 }