Clean up recipes (#596)

Makefile

@@ -40,14 +40,13 @@ evaluate:
 		fi \
 	),))
 	$(if $(filter tensor,$(PARALLEL)),export VLLM_WORKER_MULTIPROC_METHOD=spawn &&,) \
-	MODEL_ARGS="pretrained=$(MODEL),dtype=bfloat16,$(PARALLEL_ARGS),max_model_length=32768,gpu_memory_utilization=0.8,generation_parameters={max_new_tokens:32768,temperature:0.6,top_p:0.95}" && \
+	MODEL_ARGS="pretrained=$(MODEL),dtype=bfloat16,$(PARALLEL_ARGS),max_model_length=32768,max_num_batched_tokens=32768,gpu_memory_utilization=0.8,generation_parameters={max_new_tokens:32768,temperature:0.6,top_p:0.95}" && \
 	if [ "$(TASK)" = "lcb" ]; then \
 		lighteval vllm $$MODEL_ARGS "extended|lcb:codegeneration|0|0" \
 			--use-chat-template \
 			--output-dir data/evals/$(MODEL); \
 	else \
-		lighteval vllm $$MODEL_ARGS "custom|$(TASK)|0|0" \
-			--custom-tasks src/open_r1/evaluate.py \
+		lighteval vllm $$MODEL_ARGS "lighteval|$(TASK)|0|0" \
 			--use-chat-template \
 			--output-dir data/evals/$(MODEL); \
 	fi

recipes/Mistral-Small-24B-Instruct-2501/sft/config_openr1_math.yaml

@@ -1,42 +0,0 @@
-# To start the training, run the following command:
-# sbatch -N 4 --job-name=mistral_sft slurm/train.slurm Mistral-Small-24B-Instruct-2501 sft numina zero3
-
-model_name_or_path: mistralai/Mistral-Small-24B-Instruct-2501
-model_revision: main
-torch_dtype: bfloat16
-attn_implementation: flash_attention_2
-
-# Data training arguments
-# dataset_name: yentinglin/s1K-1.1-trl-format
-dataset_name: yentinglin/OpenR1-Math-220k-trl-format
-preprocessing_num_workers: 8
-
-# SFT trainer config
-bf16: true
-do_eval: true
-eval_strategy: no
-gradient_accumulation_steps: 4
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-hub_model_id: Mistral-Small-24B-Instruct-2501-Open-R1-Distill
-hub_strategy: every_save
-learning_rate: 2.0e-05
-log_level: info
-logging_steps: 1
-logging_strategy: steps
-lr_scheduler_type: cosine
-packing: true
-max_length: 32768
-max_steps: -1
-num_train_epochs: 5
-output_dir: data/Mistral-Small-24B-Instruct-2501-Open-R1-Distill
-overwrite_output_dir: true
-per_device_eval_batch_size: 1
-per_device_train_batch_size: 1
-push_to_hub: true
-report_to:
-- wandb
-save_strategy: epoch
-seed: 42
-warmup_ratio: 0.1

recipes/Qwen2.5-7B-Instruct/grpo/config_demo.yaml

@@ -1,55 +0,0 @@
-# Model arguments
-model_name_or_path: Qwen/Qwen2.5-7B-Instruct
-model_revision: main
-torch_dtype: bfloat16
-attn_implementation: flash_attention_2
-
-# Data training arguments
-dataset_name: open-r1/OpenR1-Math-cn_k12-86k
-dataset_prompt_column: problem
-system_prompt: "You are a helpful AI Assistant that provides well-reasoned and detailed responses. You first think about the reasoning process as an internal monologue and then provide the user with the answer. Respond in the following format: <think>\n...\n</think>\n<answer>\n...\n</answer>"
-
-# GRPO trainer config
-beta: 0.001
-bf16: true
-do_eval: false
-eval_strategy: "no"
-use_vllm: true
-do_eval: false
-gradient_accumulation_steps: 16
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-hub_model_id: Qwen2.5-7B-Instruct-GRPO
-hub_strategy: every_save
-learning_rate: 1.0e-06
-log_completions: true
-log_level: info
-logging_first_step: true
-logging_steps: 1
-logging_strategy: steps
-lr_scheduler_type: constant_with_warmup
-max_grad_norm: 0.2
-max_prompt_length: 1024
-max_completion_length: 4096
-max_steps: -1
-num_generations: 16
-num_train_epochs: 1
-output_dir: data/Qwen2.5-7B-Instruct-GRPO
-overwrite_output_dir: true
-per_device_train_batch_size: 4
-push_to_hub: true
-report_to:
-- wandb
-reward_funcs:
-- accuracy
-- format
-reward_weights:
-- 1.0
-- 0.2
-save_strategy: "steps"
-save_steps: 0.1
-save_total_limit: 1
-seed: 42
-temperature: 0.7
-warmup_ratio: 0.1

recipes/SmolLM2-1.7B-Instruct/sft/config.yaml

@@ -1,46 +0,0 @@
-# Model arguments
-# You can download the model and manually change the rope to 300k/500k and max_position_embeddings to 32768
-model_name_or_path: HuggingFaceTB/SmolLM2-1.7B-Instruct
-model_revision: main
-torch_dtype: bfloat16
-attn_implementation: sdpa
-
-# Data training arguments
-dataset_name: open-r1/OpenR1-Math-220k
-dataset_num_proc: 48
-
-#SFT hyperparam
-max_length: 8192 # You can set this to 32768 if you change the rope, but you need to change the config.json file 
-weight_decay: 0.0001
-optim: adamw_torch
-lr_scheduler_type: linear
-warmup_ratio: 0.1
-learning_rate: 5.0e-05
-gradient_accumulation_steps: 2
-per_device_eval_batch_size: 4
-per_device_train_batch_size: 4 # Change this depending on the context length of the model to keep a 500M GBS. 
-
-# SFT trainer config
-max_steps: -1
-num_train_epochs: 3
-bf16: true
-do_eval: false
-eval_strategy: 'no'
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-hub_model_id: OpenR1-Qwen-7B-SFT
-hub_strategy: every_save
-log_level: info
-logging_steps: 5
-logging_strategy: steps
-packing: true
-output_dir: data/OpenR1-Qwen-7B-SFT
-overwrite_output_dir: true
-push_to_hub: true
-report_to:
-- wandb
-save_strategy: "steps"
-save_steps: 500
-save_total_limit: 1
-seed: 42

recipes/SmolLM2-1.7B/sft/config.yaml

@@ -1,46 +0,0 @@
-# Model arguments
-# You can download the model and manually change the rope to 300k/500k and max_position_embeddings to 32768
-model_name_or_path: HuggingFaceTB/SmolLM2-1.7B
-model_revision: main
-torch_dtype: bfloat16
-attn_implementation: sdpa
-
-# Data training arguments
-dataset_name: open-r1/OpenR1-Math-220k 
-dataset_num_proc: 48
-
-#SFT hyperparam
-max_length: 8192 # You can set this to 32768 if you change the rope, but you need to change the config.json file 
-weight_decay: 0.0001
-optim: adamw_torch
-lr_scheduler_type: linear
-warmup_ratio: 0.1
-learning_rate: 5.0e-05
-gradient_accumulation_steps: 2
-per_device_eval_batch_size: 4
-per_device_train_batch_size: 4 # Change this depending on the context length of the model to keep a 500M GBS. 
-
-# SFT trainer config
-max_steps: -1
-num_train_epochs: 3
-bf16: true
-do_eval: false
-eval_strategy: 'no'
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-hub_model_id: OpenR1-Qwen-7B-SFT
-hub_strategy: every_save
-log_level: info
-logging_steps: 5
-logging_strategy: steps
-packing: true
-output_dir: data/OpenR1-Qwen-7B-SFT
-overwrite_output_dir: true
-push_to_hub: true
-report_to:
-- wandb
-save_strategy: "steps"
-save_steps: 500
-save_total_limit: 1
-seed: 42