-
Notifications
You must be signed in to change notification settings - Fork 1.8k
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Initial GRPO exps on the Numina dataset
- Loading branch information
1 parent
0da0f7c
commit c4d3043
Showing
21 changed files
with
1,136 additions
and
0 deletions.
There are no files selected for viewing
53 changes: 53 additions & 0 deletions
53
recipes/DeepSeek-R1-Distill-Qwen-1.5B/grpo/config_v00.00.yaml
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,53 @@ | ||
| # Model arguments | ||
| model_name_or_path: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | ||
| model_revision: main | ||
| torch_dtype: bfloat16 | ||
| attn_implementation: flash_attention_2 | ||
| # Data training arguments | ||
| dataset_name: AI-MO/NuminaMath-TIR | ||
| dataset_configs: | ||
| - all | ||
| num_processes: 7 | ||
| ddp_find_unused_parameters: false | ||
| # GRPO trainer config | ||
| use_vllm: true | ||
| vllm_device: auto | ||
| vllm_gpu_memory_utilization: 0.7 | ||
| bf16: true | ||
| do_eval: false | ||
| eval_strategy: "no" | ||
| gradient_accumulation_steps: 2 | ||
| gradient_checkpointing: true | ||
| gradient_checkpointing_kwargs: | ||
| use_reentrant: false | ||
| hub_model_id: open-r1/DeepSeek-R1-Distill-Qwen-1.5-GRPO-v00.00 | ||
| hub_strategy: every_save | ||
| learning_rate: 1.0e-06 | ||
| log_level: info | ||
| logging_steps: 1 | ||
| logging_strategy: steps | ||
| lr_scheduler_type: cosine | ||
| max_prompt_length: 1024 | ||
| max_completion_length: 4096 | ||
| max_steps: -1 | ||
| num_train_epochs: 0.1 | ||
| num_generations: 7 | ||
| output_dir: data/DeepSeek-R1-Distill-Qwen-1.5-GRPO-v00.00 | ||
| overwrite_output_dir: true | ||
| per_device_eval_batch_size: 4 | ||
| per_device_train_batch_size: 8 | ||
| push_to_hub: true | ||
| report_to: | ||
| - wandb | ||
| seed: 42 | ||
| warmup_ratio: 0.1 | ||
|
|
||
| # Saving and eval callbacks | ||
| save_strategy: "steps" | ||
| save_steps: 25 | ||
| callbacks: | ||
| - push_to_hub_revision | ||
| benchmarks: | ||
| - math_500 | ||
| - aime24 | ||
| - gpqa |
53 changes: 53 additions & 0 deletions
53
recipes/DeepSeek-R1-Distill-Qwen-1.5B/grpo/config_v00.01.yaml
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,53 @@ | ||
| # Model arguments | ||
| model_name_or_path: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | ||
| model_revision: main | ||
| torch_dtype: bfloat16 | ||
| attn_implementation: flash_attention_2 | ||
| # Data training arguments | ||
| dataset_name: AI-MO/NuminaMath-TIR | ||
| dataset_configs: | ||
| - all | ||
| num_processes: 7 | ||
| ddp_find_unused_parameters: false | ||
| # GRPO trainer config | ||
| use_vllm: true | ||
| vllm_device: auto | ||
| vllm_gpu_memory_utilization: 0.7 | ||
| bf16: true | ||
| do_eval: false | ||
| eval_strategy: "no" | ||
| gradient_accumulation_steps: 2 | ||
| gradient_checkpointing: true | ||
| gradient_checkpointing_kwargs: | ||
| use_reentrant: false | ||
| hub_model_id: open-r1/DeepSeek-R1-Distill-Qwen-1.5-GRPO-v00.01 | ||
| hub_strategy: every_save | ||
| learning_rate: 1.0e-06 | ||
| log_level: info | ||
| logging_steps: 1 | ||
| logging_strategy: steps | ||
| lr_scheduler_type: cosine | ||
| max_prompt_length: 1024 | ||
| max_completion_length: 4096 | ||
| max_steps: -1 | ||
| num_train_epochs: 0.1 | ||
| num_generations: 7 | ||
| output_dir: data/DeepSeek-R1-Distill-Qwen-1.5-GRPO-v00.01 | ||
| overwrite_output_dir: true | ||
| per_device_eval_batch_size: 4 | ||
| per_device_train_batch_size: 8 | ||
| push_to_hub: true | ||
| report_to: | ||
| - wandb | ||
| seed: 42 | ||
| warmup_ratio: 0.1 | ||
|
|
||
| # Saving and eval callbacks | ||
| save_strategy: "steps" | ||
| save_steps: 25 | ||
| callbacks: | ||
| - push_to_hub_revision | ||
| benchmarks: | ||
| - math_500 | ||
| - aime24 | ||
| - gpqa |
53 changes: 53 additions & 0 deletions
53
recipes/DeepSeek-R1-Distill-Qwen-1.5B/grpo/config_v00.02.yaml
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,53 @@ | ||
| # Model arguments | ||
| model_name_or_path: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | ||
| model_revision: main | ||
| torch_dtype: bfloat16 | ||
| attn_implementation: flash_attention_2 | ||
| # Data training arguments | ||
| dataset_name: AI-MO/NuminaMath-TIR | ||
| dataset_configs: | ||
| - all | ||
| num_processes: 7 | ||
| ddp_find_unused_parameters: false | ||
| # GRPO trainer config | ||
| use_vllm: true | ||
| vllm_device: auto | ||
| vllm_gpu_memory_utilization: 0.7 | ||
| bf16: true | ||
| do_eval: false | ||
| eval_strategy: "no" | ||
| gradient_accumulation_steps: 8 | ||
| gradient_checkpointing: true | ||
| gradient_checkpointing_kwargs: | ||
| use_reentrant: false | ||
| hub_model_id: open-r1/DeepSeek-R1-Distill-Qwen-1.5-GRPO-v00.02 | ||
| hub_strategy: every_save | ||
| learning_rate: 1.0e-06 | ||
| log_level: info | ||
| logging_steps: 1 | ||
| logging_strategy: steps | ||
| lr_scheduler_type: cosine | ||
| max_prompt_length: 1024 | ||
| max_completion_length: 4096 | ||
| max_steps: -1 | ||
| num_train_epochs: 0.1 | ||
| num_generations: 7 | ||
| output_dir: data/DeepSeek-R1-Distill-Qwen-1.5-GRPO-v00.02 | ||
| overwrite_output_dir: true | ||
| per_device_eval_batch_size: 4 | ||
| per_device_train_batch_size: 8 | ||
| push_to_hub: true | ||
| report_to: | ||
| - wandb | ||
| seed: 42 | ||
| warmup_ratio: 0.1 | ||
|
|
||
| # Saving and eval callbacks | ||
| save_strategy: "steps" | ||
| save_steps: 25 | ||
| callbacks: | ||
| - push_to_hub_revision | ||
| benchmarks: | ||
| - math_500 | ||
| - aime24 | ||
| - gpqa |
53 changes: 53 additions & 0 deletions
53
recipes/DeepSeek-R1-Distill-Qwen-1.5B/grpo/config_v00.03.yaml
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,53 @@ | ||
| # Model arguments | ||
| model_name_or_path: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | ||
| model_revision: main | ||
| torch_dtype: bfloat16 | ||
| attn_implementation: flash_attention_2 | ||
| # Data training arguments | ||
| dataset_name: AI-MO/NuminaMath-TIR | ||
| dataset_configs: | ||
| - all | ||
| num_processes: 7 | ||
| ddp_find_unused_parameters: false | ||
| # GRPO trainer config | ||
| use_vllm: true | ||
| vllm_device: auto | ||
| vllm_gpu_memory_utilization: 0.7 | ||
| bf16: true | ||
| do_eval: false | ||
| eval_strategy: "no" | ||
| gradient_accumulation_steps: 8 | ||
| gradient_checkpointing: true | ||
| gradient_checkpointing_kwargs: | ||
| use_reentrant: false | ||
| hub_model_id: open-r1/DeepSeek-R1-Distill-Qwen-1.5-GRPO-v00.03 | ||
| hub_strategy: every_save | ||
| learning_rate: 1.0e-06 | ||
| log_level: info | ||
| logging_steps: 1 | ||
| logging_strategy: steps | ||
| lr_scheduler_type: cosine | ||
| max_prompt_length: 1024 | ||
| max_completion_length: 9192 | ||
| max_steps: -1 | ||
| num_train_epochs: 0.1 | ||
| num_generations: 7 | ||
| output_dir: data/DeepSeek-R1-Distill-Qwen-1.5-GRPO-v00.03 | ||
| overwrite_output_dir: true | ||
| per_device_eval_batch_size: 4 | ||
| per_device_train_batch_size: 4 | ||
| push_to_hub: true | ||
| report_to: | ||
| - wandb | ||
| seed: 42 | ||
| warmup_ratio: 0.1 | ||
|
|
||
| # Saving and eval callbacks | ||
| save_strategy: "steps" | ||
| save_steps: 25 | ||
| callbacks: | ||
| - push_to_hub_revision | ||
| benchmarks: | ||
| - math_500 | ||
| - aime24 | ||
| - gpqa |
53 changes: 53 additions & 0 deletions
53
recipes/DeepSeek-R1-Distill-Qwen-1.5B/grpo/config_v00.04.yaml
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,53 @@ | ||
| # Model arguments | ||
| model_name_or_path: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | ||
| model_revision: main | ||
| torch_dtype: bfloat16 | ||
| attn_implementation: flash_attention_2 | ||
| # Data training arguments | ||
| dataset_name: AI-MO/NuminaMath-TIR | ||
| dataset_configs: | ||
| - all | ||
| num_processes: 7 | ||
| ddp_find_unused_parameters: false | ||
| # GRPO trainer config | ||
| use_vllm: true | ||
| vllm_device: auto | ||
| vllm_gpu_memory_utilization: 0.7 | ||
| bf16: true | ||
| do_eval: false | ||
| eval_strategy: "no" | ||
| gradient_accumulation_steps: 8 | ||
| gradient_checkpointing: true | ||
| gradient_checkpointing_kwargs: | ||
| use_reentrant: false | ||
| hub_model_id: open-r1/DeepSeek-R1-Distill-Qwen-1.5-GRPO-v00.04 | ||
| hub_strategy: every_save | ||
| learning_rate: 1.0e-06 | ||
| log_level: info | ||
| logging_steps: 1 | ||
| logging_strategy: steps | ||
| lr_scheduler_type: cosine | ||
| max_prompt_length: 1024 | ||
| max_completion_length: 16368 | ||
| max_steps: -1 | ||
| num_train_epochs: 0.1 | ||
| num_generations: 7 | ||
| output_dir: data/DeepSeek-R1-Distill-Qwen-1.5-GRPO-v00.04 | ||
| overwrite_output_dir: true | ||
| per_device_eval_batch_size: 4 | ||
| per_device_train_batch_size: 2 | ||
| push_to_hub: true | ||
| report_to: | ||
| - wandb | ||
| seed: 42 | ||
| warmup_ratio: 0.1 | ||
|
|
||
| # Saving and eval callbacks | ||
| save_strategy: "steps" | ||
| save_steps: 25 | ||
| callbacks: | ||
| - push_to_hub_revision | ||
| benchmarks: | ||
| - math_500 | ||
| - aime24 | ||
| - gpqa |
53 changes: 53 additions & 0 deletions
53
recipes/DeepSeek-R1-Distill-Qwen-1.5B/grpo/config_v00.05.yaml
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,53 @@ | ||
| # Model arguments | ||
| model_name_or_path: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | ||
| model_revision: main | ||
| torch_dtype: bfloat16 | ||
| attn_implementation: flash_attention_2 | ||
| # Data training arguments | ||
| dataset_name: AI-MO/NuminaMath-TIR | ||
| dataset_configs: | ||
| - all | ||
| num_processes: 7 | ||
| ddp_find_unused_parameters: false | ||
| # GRPO trainer config | ||
| use_vllm: true | ||
| vllm_device: auto | ||
| vllm_gpu_memory_utilization: 0.7 | ||
| bf16: true | ||
| do_eval: false | ||
| eval_strategy: "no" | ||
| gradient_accumulation_steps: 2 | ||
| gradient_checkpointing: true | ||
| gradient_checkpointing_kwargs: | ||
| use_reentrant: false | ||
| hub_model_id: open-r1/DeepSeek-R1-Distill-Qwen-1.5-GRPO-v00.05 | ||
| hub_strategy: every_save | ||
| learning_rate: 1.0e-06 | ||
| log_level: info | ||
| logging_steps: 1 | ||
| logging_strategy: steps | ||
| lr_scheduler_type: cosine | ||
| max_prompt_length: 1024 | ||
| max_completion_length: 4096 | ||
| max_steps: -1 | ||
| num_train_epochs: 0.1 | ||
| num_generations: 7 | ||
| output_dir: data/DeepSeek-R1-Distill-Qwen-1.5-GRPO-v00.05 | ||
| overwrite_output_dir: true | ||
| per_device_eval_batch_size: 4 | ||
| per_device_train_batch_size: 8 | ||
| push_to_hub: true | ||
| report_to: | ||
| - wandb | ||
| seed: 42 | ||
| warmup_ratio: 0.1 | ||
|
|
||
| # Saving and eval callbacks | ||
| save_strategy: "steps" | ||
| save_steps: 25 | ||
| callbacks: | ||
| - push_to_hub_revision | ||
| benchmarks: | ||
| - math_500 | ||
| - aime24 | ||
| - gpqa |
53 changes: 53 additions & 0 deletions
53
recipes/DeepSeek-R1-Distill-Qwen-1.5B/grpo/config_v00.06.yaml
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,53 @@ | ||
| # Model arguments | ||
| model_name_or_path: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | ||
| model_revision: main | ||
| torch_dtype: bfloat16 | ||
| attn_implementation: flash_attention_2 | ||
| # Data training arguments | ||
| dataset_name: AI-MO/NuminaMath-TIR | ||
| dataset_configs: | ||
| - all | ||
| num_processes: 7 | ||
| ddp_find_unused_parameters: false | ||
| # GRPO trainer config | ||
| use_vllm: true | ||
| vllm_device: auto | ||
| vllm_gpu_memory_utilization: 0.7 | ||
| bf16: true | ||
| do_eval: false | ||
| eval_strategy: "no" | ||
| gradient_accumulation_steps: 8 | ||
| gradient_checkpointing: true | ||
| gradient_checkpointing_kwargs: | ||
| use_reentrant: false | ||
| hub_model_id: open-r1/DeepSeek-R1-Distill-Qwen-1.5-GRPO-v00.06 | ||
| hub_strategy: every_save | ||
| learning_rate: 1.0e-06 | ||
| log_level: info | ||
| logging_steps: 1 | ||
| logging_strategy: steps | ||
| lr_scheduler_type: cosine | ||
| max_prompt_length: 1024 | ||
| max_completion_length: 4096 | ||
| max_steps: -1 | ||
| num_train_epochs: 0.1 | ||
| num_generations: 7 | ||
| output_dir: data/DeepSeek-R1-Distill-Qwen-1.5-GRPO-v00.06 | ||
| overwrite_output_dir: true | ||
| per_device_eval_batch_size: 4 | ||
| per_device_train_batch_size: 8 | ||
| push_to_hub: true | ||
| report_to: | ||
| - wandb | ||
| seed: 42 | ||
| warmup_ratio: 0.1 | ||
|
|
||
| # Saving and eval callbacks | ||
| save_strategy: "steps" | ||
| save_steps: 25 | ||
| callbacks: | ||
| - push_to_hub_revision | ||
| benchmarks: | ||
| - math_500 | ||
| - aime24 | ||
| - gpqa |
Oops, something went wrong.