The dataset viewer should be available soon. Please retry later.
model_path string | data_file string | num_epochs int64 | max_seq_length int64 | per_device_batch_size int64 | gradient_accumulation_steps int64 | fsdp string | fsdp_transformer_cls string | world_size int64 | router_aux_loss_coef float64 | output_router_logits bool | defaults dict |
|---|---|---|---|---|---|---|---|---|---|---|---|
/root/userdata/liangyan/routing_analysis/models/Marco-Nano-Base | /root/userdata/liangyan/routing_analysis/finetuning/data/splits/hau_Latn_500k.jsonl | 1 | 8,192 | 1 | 8 | full_shard auto_wrap | Qwen3MoeDecoderLayer | 8 | 0.001 | true | {
"learning_rate": 0.00001,
"warmup_ratio": 0.05,
"weight_decay": 0.1,
"adam_beta1": 0.9,
"adam_beta2": 0.95,
"max_grad_norm": 1,
"lr_scheduler_type": "cosine",
"max_seq_length": 8192,
"logging_steps": 1,
"save_steps": 500,
"save_total_limit": 2,
"seed": 42
} |
routing_analysis checkpoint archive
This public dataset repository stores checkpoint files from the
routing_analysis filesystem snapshot while preserving their original paths
under routing_analysis/.
The tree routing_analysis/finetuning/phase2_full/checkpoints/ is explicitly
excluded. All other regular files classified under checkpoint directories are
included, including small code and configuration files needed to keep those
checkpoint directories complete.
Files are uploaded individually through Hugging Face Xet; they are not packed
into tar archives. The known empty checkpoint root
paper/pretraining_documentation_token_counts/checkpoints/ is represented by
a .gitkeep placeholder. Local symlink semantics, permissions, ownership, and
timestamps are not represented by the Hub repository.
- Downloads last month
- 1,154