Instructions to use arcadia-impact/bindfn4b-ckpt with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use arcadia-impact/bindfn4b-ckpt with Transformers:
# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("arcadia-impact/bindfn4b-ckpt", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Upload lowdiv-g0/axolotl.yaml with huggingface_hub
Browse files- lowdiv-g0/axolotl.yaml +70 -0
lowdiv-g0/axolotl.yaml
ADDED
|
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
base_model: /workspace/bindfn4b_bases/g0
|
| 2 |
+
trust_remote_code: false
|
| 3 |
+
plugins:
|
| 4 |
+
- axolotl.integrations.liger.LigerPlugin
|
| 5 |
+
- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
|
| 6 |
+
liger_fused_linear_cross_entropy: true
|
| 7 |
+
liger_rope: true
|
| 8 |
+
liger_rms_norm: true
|
| 9 |
+
liger_glu_activation: true
|
| 10 |
+
adapter: lora
|
| 11 |
+
lora_r: 64
|
| 12 |
+
lora_alpha: 128
|
| 13 |
+
lora_dropout: 0.05
|
| 14 |
+
lora_target_linear: true
|
| 15 |
+
datasets:
|
| 16 |
+
- path: /workspace/scimt/data/g_rows_bindfn4b_lowdiv_g0
|
| 17 |
+
type: chat_template
|
| 18 |
+
field_messages: messages
|
| 19 |
+
eot_tokens:
|
| 20 |
+
- <end_of_turn>
|
| 21 |
+
chat_template: jinja
|
| 22 |
+
chat_template_jinja: /workspace/scimt/src/scimt/train/stages/assets/gemma3_chat_template.jinja
|
| 23 |
+
dataset_prepared_path: /workspace/bindfn4b_lowdiv/prepared_shared
|
| 24 |
+
dataset_processes: 16
|
| 25 |
+
sequence_len: 2048
|
| 26 |
+
sample_packing: false
|
| 27 |
+
pad_to_sequence_len: false
|
| 28 |
+
bf16: true
|
| 29 |
+
tf32: true
|
| 30 |
+
flash_attention: true
|
| 31 |
+
gradient_checkpointing: false
|
| 32 |
+
micro_batch_size: 32
|
| 33 |
+
gradient_accumulation_steps: 2
|
| 34 |
+
num_epochs: 10
|
| 35 |
+
max_steps: 5000
|
| 36 |
+
optimizer: adamw_torch_fused
|
| 37 |
+
learning_rate: 0.0001
|
| 38 |
+
weight_decay: 0.01
|
| 39 |
+
max_grad_norm: 1.0
|
| 40 |
+
lr_scheduler: cosine
|
| 41 |
+
cosine_min_lr_ratio: 0.1
|
| 42 |
+
warmup_steps: 100
|
| 43 |
+
train_on_inputs: false
|
| 44 |
+
logging_steps: 5
|
| 45 |
+
save_strategy: steps
|
| 46 |
+
save_steps: 100000
|
| 47 |
+
save_only_model: true
|
| 48 |
+
checkpoint_schedule:
|
| 49 |
+
- 1
|
| 50 |
+
- 3
|
| 51 |
+
- 10
|
| 52 |
+
- 30
|
| 53 |
+
- 60
|
| 54 |
+
- 100
|
| 55 |
+
- 150
|
| 56 |
+
- 200
|
| 57 |
+
- 300
|
| 58 |
+
- 450
|
| 59 |
+
- 600
|
| 60 |
+
- 900
|
| 61 |
+
- 1200
|
| 62 |
+
- 1500
|
| 63 |
+
- 2000
|
| 64 |
+
- 2500
|
| 65 |
+
- 3000
|
| 66 |
+
- 4000
|
| 67 |
+
- 5000
|
| 68 |
+
seed: 42
|
| 69 |
+
output_dir: /workspace/bindfn4b_lowdiv/g0/checkpoints
|
| 70 |
+
save_total_limit: 25
|