How to reproduce the performance on the ReDial dataset?

[dandyxxxx]

I trained according to the code provided on GitHub, but since the dataset link you provided cannot be opened, I used mapping based objects_ Lang=en_ 202112.ttl dataset. The final results of my training are as follows:

conv:
'test/dist@2': 0.310709750246931, 'test/dist@3': 0.49851841399746016, 'test/dist@4': 0.6383519119514605
rec:
'test/recall@1': 0.029324894514767934, 'test/recall@10': 0.16729957805907172, 'test/recall@50': 0.37953586497890296

(1)These results differ greatly from the results presented in the paper. Can you give me some guidance? I hope to reproduce results similar to yours. Thank you very much.
(2)According to your paper, do I need to set --n_prefix_conv 50 in the train_conv.py and --use_resp in the train_rec. py?

[linshan-79]

I also encountered the same problem as you. Since the author provided the missing files, I trained according to the guidance. However the metric results also similar to yours. Here are the detail of redial dataset metrics:

conv:

'test/dist@2': 0.26710879074361504, 'test/dist@3': 0.4199238041484408, 'test/dist@4': 0.5233526174686045,

rec:

'test/recall@1': 0.035443037974683546, 'test/recall@10': 0.1729957805907173, 'test/recall@50': 0.3744725738396624, 

Here is my config of conversational task:

accelerate launch train_conv.py \
           --dataset redial \
           --tokenizer ~/model/DialoGPT-small \
           --model ~/model/DialoGPT-small \
           --text_tokenizer ~/model/roberta-base \
           --text_encoder ~/model/roberta-base \
           --n_prefix_conv 50 \
           --prompt_encoder ${prompt_encoder_dir}/final \
           --num_train_epochs 10 \
           --gradient_accumulation_steps 1 \
           --ignore_pad_token_for_loss \
           --per_device_train_batch_size 8 \
           --per_device_eval_batch_size 16 \
           --num_warmup_steps 6345 \
           --context_max_length 200 \
           --resp_max_length 183 \
           --prompt_max_length 200 \
           --entity_max_length 32 \
           --learning_rate 1e-4 \
           --output_dir ${output_dir} \
           --log_all

(1)@dandyxxxx, you can see that I set 'n_prefix_conv=50,' but the results don't match the paper. Could you share your configuration details? Maybe we can work together to solve the problem. Thank you very much!
(2)@wxl1999 Thanks for your work! I learned a lot from your paper and code as a beginner. I'm thinking the issue might be related to the kg module not correctly capturing relations from the dataset. Could you provide some guidance? Thank you very much!

[wxl1999]

Sorry for the late reply!

• The pre-training stage is very important for the final performance. You should observe a very good performance since the answer is actually provided in the response.
  

  • If your pre-training is not so good, you cannot observe a continuous drop in the loss curve.
    

• Once your pre-training is well conducted, you will observe similar performance for the recommendation task with fine-tuning.

• As for the conversation task, since distinct is not a very reliable metric (you can observe continuous performance gain if you do not stop training), I suggest you do not pay too much attention to this, but focus more on human evaluation. This is also the practice for large language models.

• About the evaluation for conversational recommendation, you can also refer to this paper: Rethinking the Evaluation for Conversational Recommendation in the Era of Large Language Models

Hope this can help you!

[linshan-79]

Thanks for your replying! This help me a lot.

[LLLip-dev]

@linshan-79 I have the same problem. Did you finally solve it? Thank you so much.