rpDungeon/GLM-4.7-Flash-Springdragon

output-2

This model is a fine-tuned version of zai-org/GLM-4.7-Flash.

W&B run: https://wandb.ai/cooawoo-personal/huggingface/runs/ay5ml51v

Training procedure

Hyperparameters

Parameter	Value
Learning rate	2e-05
LR scheduler	SchedulerType.COSINE
Per-device batch size	1
Gradient accumulation	4
Effective batch size	4
Epochs	2
Max sequence length	8192
Optimizer	OptimizerNames.ADAMW_TORCH_FUSED
Weight decay	0.01
Warmup ratio	0.03
Max gradient norm	1.0
Precision	bf16
Gradient checkpointing	yes
Loss type	nll
Chunked cross-entropy	yes

LoRA configuration

Parameter	Value
Rank (r)	32
Alpha	16
Target modules	kv_a_proj_with_mqa, kv_b_proj, mlp.down_proj, mlp.gate_proj, mlp.up_proj, o_proj, q_a_proj, q_b_proj, shared_expert.down_proj, shared_expert.gate_proj, shared_expert.up_proj
rsLoRA	yes

Dataset statistics

Dataset	Samples	Total tokens	Trainable tokens
rpDungeon/some-revised-datasets/springdragon_processed.jsonl	2,473	5,421,492	5,421,492

Training config

model_name_or_path: zai-org/GLM-4.7-Flash
data_config: data.yaml
prepared_dataset: prepared
output_dir: output-2
attn_implementation: flash_attention_2
bf16: true
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
use_cce: true
padding_free: false
dataloader_num_workers: 4
dataloader_pin_memory: true
aux_loss_top_prob_weight: 0.05
neftune_noise_alpha: 5
max_length: 8192
per_device_train_batch_size: 1
gradient_accumulation_steps: 4
truncation_strategy: split
use_peft: true
lora_r: 32
lora_alpha: 16
lora_dropout: 0.0
use_rslora: true
lora_target_modules:
- q_a_proj
- q_b_proj
- kv_a_proj_with_mqa
- kv_b_proj
- o_proj
- shared_expert.gate_proj
- shared_expert.up_proj
- shared_expert.down_proj
- mlp.gate_proj
- mlp.up_proj
- mlp.down_proj
model_init_kwargs:
  trust_remote_code: true
  torch_dtype: bfloat16
trust_remote_code: true
optim: adamw_torch_fused
learning_rate: 2.0e-05
lr_scheduler_type: cosine
warmup_ratio: 0.03
weight_decay: 0.01
max_grad_norm: 1.0
num_train_epochs: 2
logging_steps: 1
disable_tqdm: false
saves_per_epoch: 4
eval_strategy: 'no'
save_total_limit: 3
report_to: wandb
run_name: glm47-sonic-springdragon

Data config

datasets:
- path: rpDungeon/some-revised-datasets
  data_files: springdragon_processed.jsonl
  type: text
  columns:
  - text
  truncation_strategy: split
shuffle_datasets: true
shuffle_combined: true
shuffle_seed: 42
eval_split: 0
split_seed: 42
assistant_only_loss: false

Framework versions

• PEFT 0.18.1
• Loft: 0.1.0
• Transformers: 5.3.0
• Pytorch: 2.9.1
• Datasets: 4.6.1
• Tokenizers: 0.22.2