Instruction Following on AlpacaEval 805 instructions (test)
79.91Win RateLoRA (upper bound)
Evaluation Results
| Method | Links | |
|---|---|---|
| LoRA (upper bound)Base LLM=MISTRAL-7B, Params Add=110M2024.10 | 79.91 | |
| LoRA (upper bound)Base LLM=LLAMA2-13B, Params Add=110M2024.10 | 75.16 | |
| LoRA (upper bound)Base LLM=LLAMA2-7B, Params Add=110M2024.10 | 68.18 | |
| CMCBase LLM=LLAMA2-70B, Params Add=110M, Delta Model=TINYLLAMA-110M, Template Model=LLAMA2-7B2024.10 | 49.81 | |
| CMCBase LLM=LLAMA2-13B, Params Add=110M, Delta Model=TINYLLAMA-110M, Template Model=LLAMA2-7B2024.10 | 39.04 | |
| CMCBase LLM=MISTRAL-7B, Params Add=110M, Delta Model=TINYLLAMA-110M, Template Model=LLAMA2-7B2024.10 | 33.29 | |
| CMCBase LLM=LLAMA2-7B, Params Add=110M, Delta Model=TINYLLAMA-110M, Template Model=LLAMA2-7B2024.10 | 30.41 | |
| Vanilla Base ModelBase LLM=LLAMA2-70B, Params Add=02024.10 | 11.55 | |
| Proxy-tuningBase LLM=LLAMA2-13B, Params Add=220M, Anti-expert=TINYLLAMA-110M2024.10 | 10.47 | |
| Proxy-tuningBase LLM=LLAMA2-70B, Params Add=220M, Anti-expert=TINYLLAMA-110M2024.10 | 8.59 | |
| Proxy-tuningBase LLM=LLAMA2-7B, Params Add=220M, Anti-expert=TINYLLAMA-110M2024.10 | 8.47 | |
| Vanilla Base ModelBase LLM=MISTRAL-7B, Params Add=02024.10 | 6.83 | |
| Vanilla Base ModelBase LLM=LLAMA2-13B, Params Add=02024.10 | 5.34 | |
| Vanilla Base ModelBase LLM=LLAMA2-7B, Params Add=02024.10 | 4.22 |