Training with a large json dataset (>650K) throw error:pyarrow.lib.ArrowInvalid: offset overflow while concatenating arrays

[bofei5675]

Please check that this issue hasn't been reported before.

• I searched previous Bug Reports didn't find any similar reports.

Expected Behavior

Should be able to train just like using small dataset.

Current behaviour

Throw error such as

[rank0]:   File "pyarrow/table.pxi", line 4387, in pyarrow.lib.Table.combine_chunks
[rank0]:   File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
[rank0]:   File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
[rank0]: pyarrow.lib.ArrowInvalid: offset overflow while concatenating arrays

Steps to reproduce

Prepare a dataset, configure it as

datasets:
  - path: data/dataset-test.json
    ds_type: json
    type: sharegpt
    conversation: qwen-7b-chat

dataset-test.json has more than 650K rows, and each json object is something like

{
  "id": "",
  "conversations": [
    {
       "from": "xxx", "value": "xxx"
    },
   ...
  ]
}

Delete some items in this dataset to make it less or equal to 650K rows will fix this error.

Config yaml

base_model: Qwen/Qwen2-1.5B-Instruct
trust_remote_code: true

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: data/dataset-test.json
    ds_type: json
    type: sharegpt
    conversation: qwen-7b-chat

dataset_prepared_path:
val_set_size: 0
output_dir: ./checkpoints/Qwen2-1.5B-Infinity-Instruct

sequence_len: 4096
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true

adapter: 
lora_model_dir:
lora_r: 32
lora_alpha: 64
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: project_1025
wandb_entity:
wandb_watch:
wandb_name: qwen-2-1.5B-infinity-instruct
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 8
num_epochs: 3
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.00001

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: false

warmup_steps: 40
evals_per_epoch: 4
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_limit_all_gathers: true
  fsdp_sync_module_states: true
  fsdp_offload_params: true
  fsdp_use_orig_params: false
  fsdp_cpu_ram_efficient_loading: true
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_sharding_strategy: FULL_SHARD
special_tokens:
save_total_limit: 1 # Checkpoints saved at a time

Possible solution

NA

Which Operating Systems are you using?

• Linux
• macOS
• Windows

Python Version

3.10

axolotl branch-commit

main

Acknowledgements

• My issue title is concise, descriptive, and in title casing.
• I have searched the existing issues to make sure this bug has not been reported yet.
• I am using the latest version of axolotl.
• I have provided enough information for the maintainers to reproduce and diagnose the issue.