fix: numerically unstable log-odds in ORPO loss

[Mr-Neutr0n]

Bug

The ORPO loss (OddsRatioLoss in applications/ColossalChat/coati/models/loss.py) computes log-odds using a numerically fragile pattern:

chosen_odds = chosen_logp - torch.log(-torch.exp(chosen_logp) + 1.0001)

This has two problems:

1. Biased constant: The magic value 1.0001 shifts the result away from the mathematically correct value, introducing a systematic bias into the loss.
2. NaN risk: When exp(logp) > 1.0001 (which can happen due to floating-point imprecision, especially in mixed-precision training), the argument to torch.log becomes negative, producing NaN and poisoning the training run.

The mathematically correct log-odds formula is log(p) - log(1-p) = logp - log(1 - exp(logp)).

Fix

• Clamp logp to (-inf, -eps] so that exp(logp) is strictly less than 1, preventing both division by zero and negative arguments to log.
• Replace torch.log(-torch.exp(logp) + 1.0001) with torch.log1p(-torch.exp(logp)), which is the numerically correct and unbiased way to compute log(1 - exp(logp)).

This eliminates both the NaN risk and the systematic bias from the hardcoded offset.