Panga-Azazia/reward-model-v3

Description

This model is a Reward Model trained on the RobotsMali transcription scorer dataset, where the scores were assigned by human annotators.
It predicts a continuous score between 0 and 1 for a pair (audio, text), representing how well the text matches the spoken audio.

The model can be integrated as a Reward Model within RLHF pipelines to evaluate or fine-tune ASR models based on human preference scores.

---

Model Overview

The model consists of two main encoders — one for audio and one for text — followed by a small regression head that outputs a scalar score.

---

Audio Encoder

Input: Raw waveform (16 kHz)
Feature extraction: Mel-spectrogram computed from waveform (amplitude → dB)

Parameters:

• n_fft: 1024
• n_mels: 80
• hop_length: 256
• sample_rate: 16000

---

Architecture:

• 3 × (Conv1d → BatchNorm1d → ReLU).
• Kernel size: 5, stride: 1, padding: 2.
• Channel size: 128.
• Temporal pooling: AdaptiveAvgPool1d(1).

---

Text Encoder

Input: Tokenized transcription (IDs from SentencePiece tokenizer)
Architecture:

• Embedding layer: dim = 128, vocab_size = 5000
• Bidirectional LSTM: hidden size = 128, 1 layer
• Output: mean pooling over valid tokens

---

Fusion & Regression Head

Fusion: Concatenate [audio_emb, text_emb]

Regression head:

• Linear(384 → 256) → ReLU → Dropout(0.3)
• Linear(256 → 256) → ReLU
• Linear(256 → 1) → Sigmoid

Output: Scalar ∈ [0, 1] (reward score)

---

Objective

• Loss: Mean Squared Error (MSE)
• Goal: Predict the similarity score between the spoken audio and its transcription.

---

Example Usage

Comming soon ......