Finetune Problem 微调的各种问题 #270
Replies: 42 comments 81 replies
-
|
您好！很高兴发现了CogLVM 关于微调有一下问题： |
Beta Was this translation helpful? Give feedback.
-
|
这个问题请回答下，谢谢 |
Beta Was this translation helpful? Give feedback.
-
|
确实，Option 3, finetune CogAgent/CogVLM所给的captcha数据集 |
Beta Was this translation helpful? Give feedback.
-
|
这个demo在dataset.py中把图片文件的文件名作为caption，对应简单的一问一答。我也想搞明白多轮问答的数据处理应该怎样处理喂给模型。谢谢！ |
Beta Was this translation helpful? Give feedback.
-
|
seems like it is partially answered here: (closed issues里包含了这个贴里的一些问题) |
Beta Was this translation helpful? Give feedback.
-
|
我也遇到了这个问题，自定义数据集images中的图片名称对应lables中的json名称，但是该如何喂给模型，还在研究中 |
Beta Was this translation helpful? Give feedback.
-
|
hi,我在按照示例来微调
我自己的参数是(_name_or_path='/mntnlp/common_base_model/cogvqa/cogagent', architectures=['CogAgentForCausalLM'], attention_dropout=0.1, auto_map={'AutoConfig': 'configuration_cogagent.CogAgentConfig', 'AutoModelForCausalLM': 'modeling_cogagent.CogAgentForCausalLM'}, batch_from_same_dataset=False, batch_size=4, bf16=False, block_size=10000, bos_token_id=1, checkpoint_activations=False, checkpoint_num_layers=1, checkpoint_skip_layers=0, cross_compute_hidden_size=1024, cross_hidden_size=1024, cross_image_pix=1120, cross_image_size=1120, cuda=True, deepscale=False, deepscale_config=None, deepspeed=True, deepspeed_activation_checkpointing=False, deepspeed_config={'train_micro_batch_size_per_gpu': 4, 'gradient_accumulation_steps': 1, 'gradient_clipping': 0.1, 'fp16': {'enabled': False, 'loss_scale': 0, 'loss_scale_window': 200, 'hysteresis': 2, 'min_loss_scale': 0.01}, 'bf16': {'enabled': False}, 'optimizer': {'type': 'AdamW', 'params': {'lr': 0.0001, 'weight_decay': 0.01}}}, deepspeed_mpi=False, device=0, distributed_backend='nccl', drop_path=0.0, eos_token_id=2, epochs=None, eva_args={'model_parallel_size': 1}, eval_batch_size=None, eval_interval=None, eval_iters=100, exit_interval=None, experiment_name='finetune-/mntnlp/common_base_model/cogvqa', fp16=False, from_pretrained='/mntnlp/common_base_model/cogvqa', gradient_accumulation_steps=1, hidden_act='silu', hidden_dropout=0.1, hidden_size=4096, hidden_size_per_attention_head=None, ignore_pad_token_for_loss=True, image_length=256, initializer_range=0.02, inner_hidden_size=None, input_source='interactive', intermediate_size=11008, iterable_dataset=False, layer_range=None, layernorm_epsilon=1e-05, layernorm_order='pre', length_penalty=0.0, load=None, local_rank=0, local_tokenizer='/mntnlp/common_base_model/vicuna_v1.5_7b', log_interval=50, lora_rank=50, lr=0.0001, lr_decay_iters=None, lr_decay_ratio=0.1, lr_decay_style='cosine', make_vocab_size_divisible_by=128, master_ip='127.0.0.1', master_port='16666', max_inference_batch_size=12, max_length=400, max_position_embeddings=2048, max_sequence_length=512, min_tgt_length=0, mode='finetune', model_parallel_size=1, no_load_rng=False, no_repeat_ngram_size=0, no_save_rng=False, num_attention_heads=32, num_beams=1, num_hidden_layers=32, num_layers=6, num_multi_query_heads=0, num_workers=1, out_seq_length=256, output_path='./samples', pad_token_id=0, pre_seq_len=8, prefetch_factor=4, rank=0, resume_dataloader=True, rms_norm_eps=1e-05, save=None, save_args=False, save_interval=5000, seed=1234, skip_init=False, split='1000,1,1', strict_eval=False, summary_dir='', temperature=1.0, template_version='chat', test_data=None, tie_word_embeddings=False, tokenizer_type='fake', top_k=0, top_p=0.0, torch_dtype='bfloat16', train_data=['./archive_split/train'], train_data_weights=None, train_iters=2000, transformers_version='4.36.0.dev0', use_cache=True, use_gpu_initialization=False, use_lora=True, use_ptuning=False, use_qlora=False, valid_data=['./archive_split/valid'], version='chat', vision_config={'dropout_prob': 0.0, 'hidden_act': 'gelu', 'hidden_size': 1792, 'image_size': 224, 'in_channels': 3, 'intermediate_size': 15360, 'layer_norm_eps': 1e-06, 'num_heads': 16, 'num_hidden_layers': 63, 'num_positions': 257, 'patch_size': 14}, vit_checkpoint_activations=False, vocab_size=32000, warmup=0.02, weight_decay=0.01, with_id=False, world_size=1, zero_stage=0) 这个inner_hidden_size默认怎么设置呢，感激 |
Beta Was this translation helpful? Give feedback.
-
|
解决了吗 |
Beta Was this translation helpful? Give feedback.
-
|
enable = ["encoder", "cross_attention", "linear_proj", 'mlp.vision', 'rotary.vision', 'eoi', 'boi', 'vit'] |
Beta Was this translation helpful? Give feedback.
-
|
同问，麻烦回复一下 |
Beta Was this translation helpful? Give feedback.
-
|
This can be somewhat a duplicated question, So here are my questions:
|
Beta Was this translation helpful? Give feedback.
-
|
Hmm with default setting, after lora there are still 11.6b parameters, changing the rank to 1 only decreased it to 11.5b. So you guys didn't apply it on some weights right. I would appreciate it a lot if you guys can give me some advise for modify the lora range. |
Beta Was this translation helpful? Give feedback.
-
|
我在8*3090上试图精调CogAgent模型，设置MP_SIZE=4，内存（500G）在模型加载阶段就会爆掉。请问有什么办法能够减少这部分内存占用吗？ |
Beta Was this translation helpful? Give feedback.
-
|
我发现他们lora设置的很奇怪，按lora的paper应该是lora apply在所有attention的layer上然后trainable parameter设置成只有lora出来的心得delta w。但他们把lora的parameter加到了trainable parameter里。我把这个改了加上deepspeed model parameters之外的video memory损耗就相当少了。 parameters 他们用了bfloat16 大概要35g |
Beta Was this translation helpful? Give feedback.
-
|
这个repo他们还没implement batch inference.. |
Beta Was this translation helpful? Give feedback.
-
|
感谢解答！麻烦问下具体应该在哪里修改呢？我之前做小模型2333刚刚接触LLM的Lora精调之类的…… |
Beta Was this translation helpful? Give feedback.
-
|
遇到了跟 #268 中一样的问题。除了MP_SIZE和NUM_GPUS_PER_WORKER外没有修改其他参数，基于官方SAT拖的权重进行精调，在 |
Beta Was this translation helpful? Give feedback.
-
|
这个感觉是你输入没有padding? |
Beta Was this translation helpful? Give feedback.
-
|
这个是loading checkpoint后，处理lora层的时候报的错，应该没关系？稍微复原了一下Traceback是这样的： 这个64会随着MP_SIZE的不同而变化。在此之前的logline是这样的： 顺带一提cogagent的logline从building FineTuneTrainCogAgentModel model之后就开始重复……CogVLM没有这个问题 |
Beta Was this translation helpful? Give feedback.
-
|
调试了一下还是没搞懂为什么会出错，RowParallelLinear替换在前面是正常的。但是唯独此处，在 lora2.py#L94 之前传进去[4096,256]的参数，出来得到的实例就变成[4096,64]了 |
Beta Was this translation helpful? Give feedback.
-
|
我也遇到了这个问题 |
Beta Was this translation helpful? Give feedback.
-
|
想问一下，微调能放开backend的各层参数来微调吗？如果能的话，具体怎么做？ |
Beta Was this translation helpful? Give feedback.
-
|
我按照demo的实例，下载好图片并设置好路径之后执行 bash finetune_cogvlm_lora.sh，得到下面的错误，想知道是什么原因
|
Beta Was this translation helpful? Give feedback.
-
|
sat的bug，在最新版本中修复了，先uninstall掉原来的sat，然后按照以下方式安装新的sat： |
Beta Was this translation helpful? Give feedback.
-
|
这个ERROR的地方过了，谢谢 |
Beta Was this translation helpful? Give feedback.
-
|
Hello! |
Beta Was this translation helpful? Give feedback.
-
|
请问现在有没有对hf model进行Lora或者qlora的代码呢？为什么用提供的finetune script 无法微调本地下载的hf model呢？ |
Beta Was this translation helpful? Give feedback.
-
|
Good luck |
Beta Was this translation helpful? Give feedback.
-
|
hf的微调代码由于开发组较忙一直没有时间做之后会放出一个版本 |
Beta Was this translation helpful? Give feedback.
-
|
我要是写好了是不是可以给你们提一个MR之类的？ |
Beta Was this translation helpful? Give feedback.
-
|
可以，我fork的分支已经写了，欢迎提出PR，可以参考我的失败案例 |
Beta Was this translation helpful? Give feedback.
-
|
请问一下现在的eval和finetune脚本是都仅支持sat，暂时没有提供hf的版本嘛？ |
Beta Was this translation helpful? Give feedback.
-
|
CogAgent的pretrain数据和qa格式的微调数据有计划开源吗? |
Beta Was this translation helpful? Give feedback.
-
|
暂时没有 |
Beta Was this translation helpful? Give feedback.
-
|
bash finetune_demo/finetune_cogvlm_lora.sh |
Beta Was this translation helpful? Give feedback.
-
|
看着像没安装完 |
Beta Was this translation helpful? Give feedback.
-
|
我不咋会，这个按照官方来的都没成功，能帮帮我不，感谢 |
Beta Was this translation helpful? Give feedback.
-
|
我也不是非常懂。最好把错误全文完整放上来。其它的网站能搜到这个错误，不知道会不会有帮助。 |
Beta Was this translation helpful? Give feedback.
-
|
能加个联系方式帮我看看不。球球了 |
Beta Was this translation helpful? Give feedback.
-
|
错误老长了，而且是我修改了好几个能解决的了，然后出现的，找了很久没找到，估计是我开始就有步骤不对造成的。 |
Beta Was this translation helpful? Give feedback.
-
|
finetune cogvlm最后一步save出来的文件大小是基底模型文件的两倍，有60+G，是因为float精度吗？请问怎样保存成比较小的文件可直接用于inference？ the final saved file in finetune cogvlm is almost twice as big as the base model file, 60+G, is it float precision? How to save to smaller file that can be used for inferencing? |
Beta Was this translation helpful? Give feedback.
-
|
可以手动把.pt里的除了'module'以外的状态删掉。 import torch
data = torch.load('mp_rank_00.pt')
for k in list(data.keys()):
if k != 'module':
del data[k]
torch.save(data, 'mp_rank_00_new.pt') |
Beta Was this translation helpful? Give feedback.
-
|
对于cogvlm微调，训练时采用半精度训练，模型保存时，也会保存成半精度的；但是通过对比开放出来的SAT/cogvlm-chat-v1.1/mp_rank_00_model_states.pt，与自己微调训练出来的mp_rank_00_model_states.pt，中的字段 len(origin_model["frozen_param_fragments"]) = 0
len(finetune_model["frozen_param_fragments"]) = 1156
nums = [torch.prod(torch.tensor(list(v.shape))) for k, v in finetune_model["frozen_param_fragments"].items()]
tensor(14186905088)微调训练后，保存的模型中的 |
Beta Was this translation helpful? Give feedback.
-
|
追踪到了sat中的代码，发现把 deepspeed=None，mode=‘inference’，保存出来的文件是较小的。和merge_model.py中的设置一致。（merge_model.py并没有merge，只是读取保存的文件再转成inferrence版本。好像没有必要存在） |
Beta Was this translation helpful? Give feedback.
-
|
merge_model.py是用来合并模型并行的，需要用torchrun多卡启动。如果你训练没用模型并行那就不需要了。 |
Beta Was this translation helpful? Give feedback.
-
|
明白了，谢谢解释 |
Beta Was this translation helpful? Give feedback.
-
训练过程中模型阶段性保存，可以将这部分去掉？如何去掉？ |
Beta Was this translation helpful? Give feedback.
-
|
这个不是指的merge_model过程中，而是在training_main过程中，我按照一定save_iters，来保存中间模型，以期后续具体测试选择一个表现最好的，中间过程模型保存时，我查看了下中间保存的模型，同样会将frozen_param_fragments权重进行保存，从而使得占用的存储翻倍； model, args = FineTuneTrainCogVLMModel.from_pretrained(args.from_pretrained, args, overwrite_args={'model_parallel_size': args.model_parallel_size} if args.model_parallel_size != 1 else {})
if args.use_ptuning:
model.add_mixin("ptuning", PTuningV2Mixin(args.num_layers, args.hidden_size // args.num_attention_heads, args.num_attention_heads, args.pre_seq_len))
if args.use_lora:
model.add_mixin("lora", LoraMixin(args.num_layers, args.lora_rank, layer_range=args.layer_range), reinit=True)
model.get_mixin("eva").vit_model.add_mixin("lora", LoraMixin(args.eva_args['num_layers'], args.lora_rank, layer_range=args.layer_range), reinit=True)
elif args.use_qlora:
model.add_mixin("lora", LoraMixin(args.num_layers, args.lora_rank, layer_range=args.layer_range, qlora=True), reinit=True)
if args.use_qlora and torch.cuda.is_available():
model = model.to('cuda')
from utils.utils import llama2_tokenizer
tokenizer = llama2_tokenizer(args.local_tokenizer, signal_type=args.version)
image_processor = get_image_processor(args.eva_args["image_size"][0])
text_processor = llama2_text_processor(tokenizer, args.max_length, args.image_length)
model = training_main(args, model_cls=model, forward_step_function=forward_step, create_dataset_function=partial(create_dataset_function, image_processor, text_processor), collate_fn=data_collator, forward_step_eval=forward_step_eval)
if args.use_lora:
model.get_mixin("lora").merge_lora()
model.get_mixin("eva").vit_model.get_mixin("lora").merge_lora()
args.use_lora = False
args.save = "checkpoints/merged_lora_cogvlm{}".format(args.eva_args["image_size"][0])
from sat.training.model_io import save_checkpoint
save_checkpoint(1, model, None, None, args) |
Beta Was this translation helpful? Give feedback.
-
|
好吧，问问题的人为什么换了……所以你是要解决另一个问题，前一个人想解决的是内存的问题。你想解决的是硬盘的问题…… |
Beta Was this translation helpful? Give feedback.
-
|
嗯，主要想处理training_main过程中的，中间模型保存时包含的frozen_param_fragments这部分，其他，merge_model.py, 以及合并lora权重，都可以自己手动删除， |
Beta Was this translation helpful? Give feedback.
-
|
这个没办法，可以去deepspeed仓库问一下，保存的时候他们目前是默认全保存的： |
Beta Was this translation helpful? Give feedback.
-
|
ok, 看这个api，看起来是可以有参数控制的 def _save_checkpoint(self, save_dir, tag, client_state={}, exclude_frozen_parameters=False):
save_path = self._get_ckpt_name(save_dir, tag)
zero_optimizer_state = self.zero_optimization() or self.bfloat16_enabled()
save_frozen_param = self.zero_optimization_partition_gradients() and not exclude_frozen_parameters
# A hack to save the checkpointing directory. Pipeline parallelism overrides
# module_state_dict() and uses this path to save the model. module_state_dict()
# then instead just returns None. The module_state_dict() implementation in
# PipelineEngine expects the save path to be set in self._curr_ckpt_path.
self._curr_ckpt_path = os.path.join(save_dir, tag)
module = self.module_state_dict(exclude_frozen_parameters=exclude_frozen_parameters)
self._curr_ckpt_path = None
state = dict(module=module,
buffer_names=self._get_buffer_names(),
optimizer=self.optimizer.state_dict() if self.optimizer and not zero_optimizer_state else None,
param_shapes=self._get_zero_param_shapes() if self.optimizer and zero_optimizer_state else None,
frozen_param_shapes=self._get_zero_frozen_param_attributes(self._get_param_shape_func)
if save_frozen_param else None,
shared_params=self._get_shared_params() if self.optimizer and zero_optimizer_state else None,
frozen_param_fragments=self._get_zero_frozen_param_attributes(self._get_param_fragment_func)
if save_frozen_param else None,
lr_scheduler=self.lr_scheduler.state_dict() if self.lr_scheduler is not None else None,
data_sampler=self.training_dataloader.data_sampler.state_dict() if
(self.training_dataloader is not None and self.curriculum_learning_enabled()) else None,
random_ltd=self.random_ltd_scheduler.state_dict() if self.random_ltd_enabled() else None,
sparse_tensor_module_names=self.sparse_tensor_module_names,
skipped_steps=self.skipped_steps,
global_steps=self.global_steps,
global_samples=self.global_samples,
dp_world_size=self.seq_dp_world_size,
mp_world_size=self.mp_world_size,
ds_config=self.config,
ds_version=version)
state.update(client_state) |
Beta Was this translation helpful? Give feedback.
-
|
遇到问题：使用finetune_cogvlm_demo.py 跑几张图，跑了100步，发现loss一直是0，validation pred 文字也一直没有任何变化。 训练参数在原基础上改了batch_size=1，MP_num=1, 其它都没碰。因为vram不够，可训练参数没有包含lora，仅包含vit mlp 和 ptuning。 会是什么问题呢？ |
Beta Was this translation helpful? Give feedback.
-
|
好像loss没有被计算，validation主要看实际输出的文字内容主观判断训练进程。将训练模式从base改成vqa，打印出来每一步的loss和validation pred，训练确实在进行。 |
Beta Was this translation helpful? Give feedback.
-
Beta Was this translation helpful? Give feedback.
-
|
请问怎么基于Lora微调的模型再进行Lora微调。我已经基于base模型lora微调了一个模型并保存了CKPT，我想加载这个CKPT后并联上Lora层，再在其他数据集上做微调，请问如何配置？ |
Beta Was this translation helpful? Give feedback.
-
|
运行微调sh文件的时候出现 |
Beta Was this translation helpful? Give feedback.
-
|
不要用 pip 安装 torch, 用 conda 安装二进制兼容会更好 |
Beta Was this translation helpful? Give feedback.
-
|
新建了个环境，conda 安装了torch2.1.1还是不太行..... |
Beta Was this translation helpful? Give feedback.
-
|
问题应该解决了，是cuda版本的问题，切到跟torch版本一样的cuda就解决了，但是报错显存不够。。。。。 |
Beta Was this translation helpful? Give feedback.
-
|
8卡A100运行finetune_cogagent_lora.sh爆显存.... |
Beta Was this translation helpful? Give feedback.
-
|
调成1也不行.... |
Beta Was this translation helpful? Give feedback.
-
|
NUM_GPUS_PER_WORKER=8，MP_SIZE=8 应该是可以的，注意调整 batch_size 的大小，另外内存也要保证够大 |
Beta Was this translation helpful? Give feedback.
-
|
NUM_GPUS_PER_WORKER=4，MP_SIZE=4 在 4 * A100 80G， batch_size 试过4，2都不行，最后只能调成1跑起来了 |
Beta Was this translation helpful? Give feedback.
-
|
输入文件的分辨率会对VRAM使用率有影响么？ |
Beta Was this translation helpful? Give feedback.
-
|
@ShChen233 我的 cuda 环境是11.8, 如果通过原始的 pip install 安装 requirements 的话，执行的时候会有如下告警 字面上看起来版本不匹配会影响到内存使用的效率，降级XFORMERS到cuda11.8的版本就可以解决了，不过这里得小心它会安装 2.1.0 的torch, 要在 xformer装完后再用conda 把torch装回2.1.2 如果直接用官方pytorch 2.1.2的 cuda12.1版本应该就不会有这么多问题了。 |
Beta Was this translation helpful? Give feedback.
-
|
想问一下有人试过 ，quant4 量化微调CogAgent嘛？我报维度错误：File "/root/miniconda3/envs/CogVLM/lib/python3.10/site-packages/sat/model/finetune/lora2.py", line 97, in init |
Beta Was this translation helpful? Give feedback.
-
|
能否说明一下这些参数具体是指哪些模块：enable = ["encoder", "cross_attention", "linear_proj", 'mlp.vision', 'rotary.vision', 'eoi', 'boi', 'vit']，如果要冻结是不是仅仅在这里设置就可以了 |
Beta Was this translation helpful? Give feedback.
-
|
我微调的时候，直接置空了，enable=[]；我只微调transformers的32层 |
Beta Was this translation helpful? Give feedback.
-
|
微调如果要从没有merge过的checkpoint应该如何继续呢？ |
Beta Was this translation helpful? Give feedback.
-
|
我试了，可以在推理的时候所有的 from pretrain里的AutoModel类改为 FinetuneTestModel类 |
Beta Was this translation helpful? Give feedback.
-
|
不是指用 checkpoint 推理，是指意外中断后重新提交作业如何从还没有merged过的checkpoint 继续 |
Beta Was this translation helpful? Give feedback.
-
|
我看了几个 sat 的示例尝试用 ValueError: could not find the metadata file checkpoints/finetune-cogagent-vqa-04-10-13-37/1200/latest, please check --load
raise ValueError('could not find the metadata file {}, please check --load'.format(
raise ValueError('could not find the metadata file {}, please check --load'.format(
ValueError: ValueErrorcould not find the metadata file checkpoints/finetune-cogagent-vqa-04-10-13-37/1200/latest, please check --load:
could not find the metadata file checkpoints/finetune-cogagent-vqa-04-10-13-37/1200/latest, please check --load
raise ValueError('could not find the metadata file {}, please check --load'.format(
ValueError: could not find the metadata file checkpoints/finetune-cogagent-vqa-04-10-13-37/1200/latest, please check --load |
Beta Was this translation helpful? Give feedback.
-
|
请问多轮对话如何把多个问答拼接起来？如何修改dataset.py呢？
|
Beta Was this translation helpful? Give feedback.
-
|
这个问题需要自己修改CAPTCHA，然后修改process_text内部的逻辑，自己实现。 |
Beta Was this translation helpful? Give feedback.
-
|
请问 finetune cogagent _demo 代码中使用的data_collator 返回的信息如下： 但是我看cogagent模型的forward 方法中接受的参数并不是这些，好奇中间还有哪里做了什么处理吗？
|
Beta Was this translation helpful? Give feedback.
-
|
请问如何把4卡并行lora微调出来的模型合并成一个模型。 我加载了开源的cogvlm-base-490模型，使用MP_SIZE=4做了lora微调，保存了一个CKPT，包含4个文件（mp_rank_00_model_states.pt到mp_rank_03_model_states.pt ）。请问我如何把模型合并成MP_SIZE=1的模型并存储到一个文件呢。谢谢 |
Beta Was this translation helpful? Give feedback.
-
|
请问单机2*8张V100（32G）去lora微调CogAgent会out of memory，这个是正常的吗（batch size已设置为1，mp_size=16，NUM_GPUS_PER_WORKER=8）？是否是因为数据并行+单张卡内存太小，所以导致的out of memory呢？ 另外16个V100微调时输出的日志显示单张卡上的参数量为1.5B，但rank 0这个卡上的参数量又是总参数量18B，这是为什么呢？ |
Beta Was this translation helpful? Give feedback.
-
|
I have a problem but no idea. Here is the log. (cogvlm) dhu_mbzhao_1@deeplearning-v191204-deeplearn: Could someone help me? |
Beta Was this translation helpful? Give feedback.
-
It failed to load your dataset so the |
Beta Was this translation helpful? Give feedback.
-
|
OK, I will check it again. Hope to useful! Thank you. |
Beta Was this translation helpful? Give feedback.
-
|
i've had this issue as well，it should be change your dataset or modify the dataset.py |
Beta Was this translation helpful? Give feedback.
-
|
what is this mean? Is this log impact the performance? Keyword arguments {'add_special_tokens': False} not recognized. |
Beta Was this translation helpful? Give feedback.
-
|
My test results are as follows, is this considered good or bad? [2024-07-19 09:38:21,589] [INFO] [RANK 0] validation loss at the end of training for test data | loss: 0.000000E+00 | PPL: 1.000000E+00 acc 5.319865E-02 | acc_w/o_case 5.319865E-02 | |
Beta Was this translation helpful? Give feedback.
-
|
The training script is based on deepspeed, you can config deepspeed to use wandb or tensorboard to track and visualize your experiment. |
Beta Was this translation helpful? Give feedback.
-
|
微调Cogagent的时候发现模型eval的过程中输出的pred是模型的回答，但是相对应的label却是问的问题，这是正常的嘛 |
Beta Was this translation helpful? Give feedback.
-
|
当我使用我自己构建的数据集，迭代一千次后loss很低，但是实际对话效果很差，经常出现错误的检测，是数据集的问题还是超参数设置的问题，用的base-490的模型进行lora微调 |
Beta Was this translation helpful? Give feedback.
and others
-
Please articulate any questions about model fine-tuning here; these questions will be answered by community members and officials in their spare time.
在这里阐述任何关于模型微调的问题，这些问题将由社区成员和官方在空闲的时间进行回答
Beta Was this translation helpful? Give feedback.
All reactions