Qwen2.5-3B-DAPO-math-reasoning

This model is a fine-tuned version of Qwen2.5-3B using DAPO (Decoupled Clip and Dynamic Sampling Policy Optimization) without KL penalty for mathematical reasoning.

Trained with PipelineRL.

Training Details

Datasets

Split	Datasets
Train	gsm8k_train, math_train
Test	gsm8k_test, math_500

RL Algorithm

Parameter	Value
Algorithm	DAPO (Decoupled Clip and Dynamic Sampling Policy Optimization)
Advantage Baseline	Group mean reward
Extra Inference	None
Group Structure	Required
Policy Loss	ppo
KL Coefficient	0.0
Epsilon (clip)	0.2
Discount Factor (gamma)	1.0
Divide Advantage by Std	False
Filter Zero Advantage Groups	True
Rollouts per Problem	16

DAPO extends GRPO with clip-higher (asymmetric PPO clipping), dynamic sampling (filtering zero-variance groups), token-level loss aggregation, and overlong reward shaping.

Training Hyperparameters

Parameter	Value
Base Model	Qwen/Qwen2.5-3B
Learning Rate	1e-06
LR Scheduler	cosine
Warmup Steps	25
Max Training Steps	1500
Micro Batch Size	2
Gradient Accumulation	128
Effective Batch Size	256
Sequence Length	8192
Gradient Clipping	0.3
Weight Decay	0.01
Optimizer	adamw_torch
Precision	bf16
DeepSpeed	ZeRO Stage 3

Evaluation Results

Pass@k on math reasoning benchmarks (N=32 samples per problem, temperature=1.0):

Dataset	pass@1	pass@2	pass@4	pass@8	pass@16	pass@32
GSM8K (test)	86.52	91.04	93.73	95.52	96.73	97.50
MATH-500	70.66	77.91	83.26	87.27	90.10	92.00
Overall	82.16	87.43	90.85	93.25	94.90	95.99

GSM8K test: 1319 problems · MATH-500: 500 problems · Overall: 1819 problems (overall weighted by problem count).

Training Curves

W&B Run

Full training logs: https://wandb.ai/jaygala24-team/rl-post-training/runs/qwen2.5_3b_dapo_no_kl_3a1f_4xh100_235923_finetune_72e237c1

Usage

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("jaygala24/Qwen2.5-3B-DAPO-math-reasoning", revision="step-0200")  # optional branch, e.g. "step-0400"
tokenizer = AutoTokenizer.from_pretrained("jaygala24/Qwen2.5-3B-DAPO-math-reasoning", revision="step-0200")

prompt = "Please reason step by step, and put your final answer within \\boxed{}.\n\nWhat is the sum of 123 and 456?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=4096, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

vLLM

from vllm import LLM, SamplingParams

llm = LLM(model="jaygala24/Qwen2.5-3B-DAPO-math-reasoning", revision="step-0200")  # optional branch, e.g. "step-0400"
sampling_params = SamplingParams(temperature=0.7, max_tokens=4096)

prompt = "Please reason step by step, and put your final answer within \boxed{}.

What is the sum of 123 and 456?"
outputs = llm.generate([prompt], sampling_params)
print(outputs[0].outputs[0].text)

Framework

• PipelineRL
• Transformers
• DeepSpeed (ZeRO Stage 3)