payelb/HHRLHF_TinyLlama-1.1B_aligned_with_MARS_deberta_RM
Base model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
Alignment dataset: Anthropic/hh-rlhf
Reward model: payelb/HHRLHF_reward-model-deberta-v3-base_1k_fixed_MARS
Method: PPO alignment with LoRA adapters.
Notes:
- Reward normalization and clipping enabled
- KL control enabled
- pad_token_id/eos_token_id explicitly set
- DeBERTa RM loaded on a single device (no device_map='auto')
