Mingdao Liu comments

Results 9 comments of


                                            Mingdao Liu

WARNING: tokenization mismatch 139 vs. 141

I also encountered this problem (i.e. total_len == curr_len + 2). Further debugging reveals that this is caused by a mis-tokenization of ``. Missing `tokenizer_config.json` and `special_tokens_map.json` in the converted...

[Help] 请问如何实现多卡微调？

See https://github.com/THUDM/ChatGLM-6B/tree/main/ptuning#finetune （？

貌似hotpotqa测试脚本跑不起来？

是否可以提供更详细的运行过程及报错？便于我们复现问题并寻找可能的原因

基于fastchat部署，推理异常

我们之前没有在 FastChat 上进行过推理实验。能否给出更详细的报错、不符合预期的场景或者配置？

if it is possible to conduct RLHF from env

We haven't integrated RLHF methods into AgentTuning yet and we won't be releasing related experimental results recently. I believe that would be an awesome thing to try out.

可以给个简单点的工具调用示例吗

比较简单的方式应该就是直接通过 TGI 部署 checkpoint，然后按照 llama-2-chat 的格式发送请求进行交互。由于是多任务训练的模型，工具调用指令没有特殊的要求。如果是要在论文中的 benchmark 上验证结果，可以参考 [README](https://github.com/THUDM/AgentTuning#evaluation) 中的说明。

AgentTuning 7b evaluate in HH， not expect as paper result

> How much epoch have you trained? The models are trained for 2k steps, batch size 64, sequence length 4096 with packing.

Where do the models go?

I found a simple solution to place the model ```bash # in the base directory of the repo mkdir -p models/speech_encoder cd models/speech_encoder wget https://openaipublic.azureedge.net/main/whisper/models/e5b1a55b89c1367dacf97e3e19bfd829a01529dbfdeefa8caeb59b3f1b81dadb/large-v3.pt ```

RuntimeError: The size of tensor a (5002) must match the size of tensor b (2) at non-singleton dimension 3

I have also encountered this problem today.