NewEden/GLM-Magnum-V5-Ckpts

See axolotl config

axolotl version: 0.9.0

base_model: THUDM/GLM-4-32B-0414
#base_model_ignore_patterns: "*/*"
# optionally might have model_type or tokenizer_type
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: anthracite-core/magnum-v5-sft-proto-glm4-instruct-rev1
    ds_type: parquet
    type:
shuffle_merged_datasets: true
dataset_prepared_path: ./data/magnum-32b-data
val_set_size: 0.0025
output_dir: ./data/32b-lora-out

plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
#liger_rope: false
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: true
cut_cross_entropy: true

sequence_len: 32768
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

adapter: lora
lora_model_dir:
lora_r: 128
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
peft_use_rslora: true
lora_modules_to_save:
  - embed_tokens
  - lm_head

wandb_project: 32b-magnum-lora
wandb_entity:
wandb_watch:
wandb_name: run7-lora-0.02-clip-4.5e-5
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 1
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: rex
learning_rate: 4.5e-5
max_grad_norm: 0.02

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

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
s2_attention:

warmup_steps: 40
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens:
saves_per_epoch: 2
debug:
deepspeed: ./deepspeed_configs/zero3_bf16.json
weight_decay: 0.0025
fsdp:
fsdp_config:
special_tokens:

data/32b-lora-out

This model is a fine-tuned version of THUDM/GLM-4-32B-0414 on the anthracite-core/magnum-v5-sft-proto-glm4-instruct-rev1 dataset. It achieves the following results on the evaluation set:

• Loss: 1.1206

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 4.5e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• total_eval_batch_size: 8
• optimizer: Use adamw_bnb_8bit with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 40
• num_epochs: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss
1.348	0.0024	1	1.3957
1.1269	0.2506	104	1.1851
1.1548	0.5012	208	1.1661
1.0896	0.7518	312	1.1531
1.0712	1.0024	416	1.1447
1.0552	1.2530	520	1.1362
1.0567	1.5036	624	1.1275
1.0888	1.7542	728	1.1206

Framework versions

• PEFT 0.15.2
• Transformers 4.51.3
• Pytorch 2.6.0+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1