Modellmått och justeringar

Reinforcement Learning from Human Feedback (RLHF)

Mina Parham

AI Engineer

Varför använda en referensmodell?

  • Meningslösa utdata

En serie emojis.

Reinforcement Learning from Human Feedback (RLHF)

Kontrollera modellens utdata

Diagram över RLHF-processen med kontroll av svaret.

Reinforcement Learning from Human Feedback (RLHF)

Lösning: KL-divergens

Diagram över RLHF-processen med KL-divergens.

Reinforcement Learning from Human Feedback (RLHF)

Lösning: KL-divergens

  • Straff läggs till i belöningsmodellen
  • Straffet omdirigerar modellen om den producerar irrelevanta utdata
  • KL-divergens jämför aktuell modell med belöningsmodellen

En ikon som föreställer en våg.

  • Mellan 0 och 10, och aldrig negativt
Reinforcement Learning from Human Feedback (RLHF)

Justera parametrar

generation_kwargs = {

    "min_length": -1, # don't ignore the EOS token

    "top_k": 0.0, # no top-k sampling

    "top_p": 1.0, 
    "do_sample": True, 
    "pad_token_id": tokenizer.eos_token_id, 
    "max_new_tokens": 32}

 

  • Parametrarna skickas till policymodellen
Reinforcement Learning from Human Feedback (RLHF)

Kontrollera belöningsmodellen

  • Kontrollera belöningsmodellen

  • Kontrollera utdata (belöning)

reward_model_results.head()

|ID | Comment                                     |Sentiment |Reward|
|---|---------------------------------------------|----------|------|
| 1 | This event was lit! So much fun!            | Positive |  0.9 |
| 2 | Terrible experience, never attending again. | Negative | -0.8 |
| 3 | It was okay, nothing extraordinary.         | Neutral  |  0.2 |
| 4 | The event was poorly organized and chaotic. | Negative | -0.85|
| 5 | Had an amazing time with great people!      | Positive |  0.95|
Reinforcement Learning from Human Feedback (RLHF)

Kontrollera belöningsmodellen

  • 👍 👎 Kontrollera extremfall
    extreme_positive = reward_model_results[reward_model_results['Reward'] >= 0.9]
extreme_negative = reward_model_results[reward_model_results['Reward'] <= -0.8]

  • 🧘 Säkerställ balanserad datamängd

    sentiment_distribution = reward_model_results['Sentiment'].value_counts()

  • 📊 Normalisera belöningsmodellen

    from sklearn.preprocessing import MinMaxScaler
scaler = MinMaxScaler(feature_range=(-1, 1))
scaler.fit_transform(reward_model_results[['Reward']])
Reinforcement Learning from Human Feedback (RLHF)

Nu kör vi en övning!

Reinforcement Learning from Human Feedback (RLHF)

Preparing Video For Download...