Model metrics and adjustments

Reinforcement Learning from Human Feedback (RLHF)

Mina Parham

AI Engineer

Why use a reference model?

Meaningless outputs

A series of emojis.

Checking model output

Diagram of the RLHF process with checking of the response.

Solution: KL divergence

Diagram of the RLHF process with KL divergence.

Solution: KL divergence

Penalty is added to the reward model
Penalty redirects the model if it produces unrelated outputs
KL divergence compares current and reward model

An icon representing a scale.

Between 0 and 10, and never negative

Adjusting parameters

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}

Parameters are passed to the policy model

Checking the reward model

Checking the reward model
Checking output (reward)

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|

Checking the reward model

👍 👎 Check extreme cases

extreme_positive = reward_model_results[reward_model_results['Reward'] >= 0.9]
extreme_negative = reward_model_results[reward_model_results['Reward'] <= -0.8]

🧘 Ensure balanced dataset

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

📊 Normalize the reward model

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

Let's practice!

Reinforcement Learning from Human Feedback (RLHF)