Unable to directly repro T5 results

ILQL sentiments reproduction from `main`

* PPO sentiments reproduction from `main` (gpt2) * Summarize dailymail/cnn reproduction from `main` (google/flan-t5-small)`

Report for the following PR: https://github.com/CarperAI/trlx/pull/133

Report for the following PR: https://github.com/CarperAI/trlx/pull/110

Results for the ILQL Sentiment task with LORA to observe training dynamics and memory saving.

Results for the PPO Sentiment task with LORA to observe training dynamics and memory saving.

Results for PPO/ILQL sentiments examples to demonstrate no regressions.

PPO Sentiments Benchmark on Multi-Node DDP setup

Hydra GPT-J with all but one layer unfrozen in the base backbone

Hydra GPT-J with all but one layer unfrozen in the base backbone

ILQL sentiment results for the `Add GPTNeo support` PR.

PPO sentiment results for the `Add GPTNeo support` PR.

ILQL Benchmark results for the wider Causal LM support PR.

PPO Benchmark results for the wider Causal LM support PR.