Instructions to use jaygala24/Qwen3-4B-RLOO-math-reasoning with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use jaygala24/Qwen3-4B-RLOO-math-reasoning with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="jaygala24/Qwen3-4B-RLOO-math-reasoning")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("jaygala24/Qwen3-4B-RLOO-math-reasoning")
model = AutoModelForCausalLM.from_pretrained("jaygala24/Qwen3-4B-RLOO-math-reasoning")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use jaygala24/Qwen3-4B-RLOO-math-reasoning with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "jaygala24/Qwen3-4B-RLOO-math-reasoning"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jaygala24/Qwen3-4B-RLOO-math-reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/jaygala24/Qwen3-4B-RLOO-math-reasoning

SGLang

How to use jaygala24/Qwen3-4B-RLOO-math-reasoning with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "jaygala24/Qwen3-4B-RLOO-math-reasoning" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jaygala24/Qwen3-4B-RLOO-math-reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "jaygala24/Qwen3-4B-RLOO-math-reasoning" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jaygala24/Qwen3-4B-RLOO-math-reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use jaygala24/Qwen3-4B-RLOO-math-reasoning with Docker Model Runner:
```
docker model run hf.co/jaygala24/Qwen3-4B-RLOO-math-reasoning
```

jaygala24 commited on Apr 26

Commit

45de45f

verified ·

1 Parent(s): 14c524c

Add training_config.yaml

Browse files

Files changed (1) hide show

training_config.yaml +287 -0

training_config.yaml ADDED Viewed

	@@ -0,0 +1,287 @@

+finetune:
+  data: null
+  model_class: causal-language-modeling
+  config_name: ${..model_path}
+  optim: adamw_torch
+  load_as_bf16: true
+  fp32_lm_head: ${..fp32_lm_head}
+  fp32_layer_prefix: ${..fp32_layer_prefix}
+  use_flash_attention: true
+  attn_implementation: flash_attention_2
+  auto_device_map: false
+  lora:
+    enabled: false
+    task_type: CAUSAL_LM
+    base_model_8bit: false
+    base_model_4bit: false
+    r: 16
+    alpha: 16
+    dropout: 0.05
+    bias: none
+    target_modules: []
+  force_restart: ${..force_restart}
+  resume_dataloader: false
+  train_batch_size: 2
+  valid_batch_size: 4
+  weight_decay: 0.01
+  learning_rate: 1.0e-06
+  gradient_clipping_threshold: 0.3
+  lr_scheduler_type: cosine
+  num_warmup_steps: 25
+  gradient_accumulation_passes: 128
+  gradient_checkpointing: true
+  reentrant_checkpointing: false
+  max_train_steps: 1500
+  interrupt_train_steps: -1
+  max_eval_steps: -1
+  seq_length: 8192
+  seq_packing: true
+  output_dir: ${..output_dir}/finetune
+  seed: ${..seed}
+  save_checkpoint_steps: 100
+  keep_intermediate_checkpoints: true
+  trust_remote_code: false
+  cuda_empty_cache: true
+  sft_config_name: null
+  n_examples: 0
+  log_each_n_steps: 1
+  also_save_steps: []
+  use_safetensors: true
+  save_final_training_state: true
+  seq_parallel: 1
+  objective: rl
+  input: training_data
+  send_weight_updates: true
+  queue_size: 32
+  max_lag: null
+  weight_update_interval: 1
+  pop_old_data: ${..pop_old_data}
+  attempts: 8
+  eval_callback:
+    _target_: pipelinerl.finetune.utils.dummy_eval_callback
+    config_name: ''
+  rl:
+    policy_loss: reinforce
+    divide_advantage_by_std: false
+    kl_coef: 0.0
+    final_kl_coef: 0.0
+    entropy_bonus: 0.0
+    reward_minus_kl_coef: 0.0
+    epsilon_low: 0.02
+    epsilon_high: 0.02
+    use_advantages: true
+    relu_log_p_weights: false
+    clamp_log_ratio_ref_new_value: 5
+    temperature: ${...llm.parameters.temperature}
+    aggregate_loss: sum
+    overlong_filtering: false
+    adv_estimator: rloo
+    filter_zero_advantage_groups: false
+rewards:
+  correct_answer_finished: 1.0
+  correct_answer_not_finished: 1.0
+  wrong_answer_finished: 0
+  wrong_answer_not_finished: 0
+  no_answer_finished: 0
+  no_answer_not_finished: 0
+  unparsable_finished: 0
+  unparsable_not_finished: 0
+streams:
+  backend: files
+seed: 42
+fp32_lm_head: false
+fp32_layer_prefix: lm_head
+actor:
+  log_each_n_secs: 0
+  llm_max_rollouts: 256
+  rollout_workers: 1
+  discount_factor: 1
+  problem_queue_size: 256
+  result_queue_size: 256
+  throughput_window_size: 50
+  shared_memory_entry_size: 10000000
+  rollout_policy: pipelinerl.domains.math.generate_math_rollout
+  system_prompt: Please reason step by step, and put your final answer within \boxed{}.
+  task_template: '{task}'
+  task_prompt: ''
+environment: null
+preprocess:
+  input: actor
+  output: training_data
+  n_workers: 8
+  chunk_n_groups: 2
+  raw_queue_size: 8
+  input_queue_size: 32
+  output_queue_size: 32
+  dataset_buffer_size: 0
+  ring_buffer_size: 128
+  max_ready_samples_per_lead: 64
+  pop_old_data: ${..pop_old_data}
+  shared_memory_entry_size: 100000000
+  log_every_n_samples: 128
+llm:
+  parameters:
+    max_tokens: 4096
+    temperature: 1.0
+test_llm:
+  parameters:
+    max_tokens: 4096
+    temperature: 1.0
+    top_p: 0.95
+    top_k: 50
+vllm_config:
+  use_v1: false
+  quantization: null
+  vllm_kwargs:
+    dtype: bfloat16
+    gpu-memory-utilization: 0.92
+    max-num-seqs: 64
+    max-num-batched-tokens: 16384
+    enable-chunked-prefill: ''
+    return-tokens-as-token-ids: ''
+    tensor-parallel-size: 1
+    pipeline-parallel-size: 1
+    generation-config: vllm
+    max_model_len: 8192
+    num-scheduler-steps: 8
+    disable-log-requests: ''
+    disable-frontend-multiprocessing: ''
+world:
+  replicas: 1
+  actor_fraction: 6
+  preprocessor_fraction: 0
+  finetune_fraction: 2
+  env_replicas: 1
+  actor_group_port: 9000
+  environment_start_port: 7777
+jobs:
+- kind: actor_llm
+  idx: 0
+  replica_idx: 0
+  local_idx: 0
+  node_rank: 0
+  hostname: localhost
+  port: 8080
+  gpus:
+  - 0
+  url: http://localhost:8080
+  environment_key: null
+  environment_index: null
+- kind: actor_llm
+  idx: 1
+  replica_idx: 1
+  local_idx: 1
+  node_rank: 0
+  hostname: localhost
+  port: 8081
+  gpus:
+  - 1
+  url: http://localhost:8081
+  environment_key: null
+  environment_index: null
+- kind: actor_llm
+  idx: 2
+  replica_idx: 2
+  local_idx: 2
+  node_rank: 0
+  hostname: localhost
+  port: 8082
+  gpus:
+  - 2
+  url: http://localhost:8082
+  environment_key: null
+  environment_index: null
+- kind: actor
+  idx: 3
+  replica_idx: 0
+  local_idx: 0
+  node_rank: 0
+  hostname: localhost
+  port: null
+  gpus: []
+  url: ''
+  environment_key: null
+  environment_index: null
+- kind: preprocessor
+  idx: 4
+  replica_idx: 0
+  local_idx: 0
+  node_rank: 0
+  hostname: localhost
+  port: null
+  gpus: []
+  url: ''
+  environment_key: null
+  environment_index: null
+- kind: environment
+  idx: 5
+  replica_idx: 0
+  local_idx: 0
+  node_rank: 0
+  hostname: localhost
+  port: 7777
+  gpus: []
+  url: ''
+  environment_key: math
+  environment_index: 0
+- kind: finetune
+  idx: 6
+  replica_idx: 0
+  local_idx: 0
+  node_rank: 0
+  hostname: localhost
+  port: null
+  gpus:
+  - 3
+  url: ''
+  environment_key: null
+  environment_index: null
+eval_every_n_versions: 78000
+model_path: Qwen/Qwen3-4B
+accelerate_config: null
+use_deepspeed: true
+deepspeed_config: deepspeed_stage3_bf16
+use_fsdp: false
+fsdp:
+  param_dtype: fp32
+  reduce_dtype: fp32
+  buffer_dtype: fp32
+output_dir: results/qwen3_4b_rloo_no_kl_3a1f_4xh100_236660
+force_restart: false
+pop_old_data: true
+max_lag: null
+attempts: 16
+train_subset: null
+debug:
+  mode: ''
+  streams_from: null
+  place_inference_workers: true
+  use_existing_llms: false
+me:
+  job_idx: null
+wandb:
+  use_wandb: true
+  fail_on_init_error: false
+  init_timeout: 120
+  wandb_id: null
+  wandb_name: null
+  wandb_entity_name: jaygala24-team
+  wandb_project_name: rl-post-training
+  wandb_resume: always
+  wandb_use_basename: true
+  wandb_workspace_root: results
+  wandb_group: qwen3_4b_rloo_no_kl_3a1f_4xh100_236660
+  wandb_dir: null
+  tags: []
+environments:
+- key: math
+  mode: remote
+  _target_: pipelinerl.domains.math.MathEnvironment
+environment_key: math
+dataset_loader: pipelinerl.domains.math.load_datasets
+train_dataset_names:
+- gsm8k_train
+- math_train
+test_dataset_names:
+- gsm8k_test
+- math_500