gpt-oss-20b-multilingual-reasoning

23
20.0B
license:apache-2.0
by
ZeroAgency
Language Model
OTHER
20B params
New
23 downloads
Early-stage
Try on Hugging FaceAdd to Compare
Edge AI:
Mobile
Laptop
Server
45GB+ RAM
Mobile
Laptop
Server
Quick Summary

This model is a fine-tuned version of axolotl-ai-co/gpt-oss-20b-dequantized on the HuggingFaceH4/Multilingual-Thinking dataset.

Device Compatibility

Mobile
4-6GB RAM
Laptop
16GB RAM
Server
GPU
Minimum Recommended
19GB+ RAM

Code Examples

the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true
the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loadingyamlvllm
# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
base_model: axolotl-ai-co/gpt-oss-20b-dequantized

use_kernels: false

dp_shard_size: 8  # requires 2x8xH100 nodes

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-20b/
#save_only_model: true

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

wandb_project: gpt-oss-20b
wandb_name: fft-20b

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1

optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
lr_scheduler: constant_with_warmup
learning_rate: 2e-5
load_best_model_at_end: false

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.03

special_tokens:
eot_tokens:
  - "<|end|>"

#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
fsdp_version: 2
fsdp_config:
  offload_params: true
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true

Deploy This Model

Production-ready deployment in minutes

Together.ai

Instant API access to this model

Fastest API

Production-ready inference API. Start free, scale to millions.

Try Free API

Replicate

One-click model deployment

Easiest Setup

Run models in the cloud with simple API. No DevOps required.

Deploy Now

Disclosure: We may earn a commission from these partners. This helps keep LLMYourWay free.