Qwen3.5-4B · 8k-token budget · grpo-overlong-penalty · awareshrimp
RL-finetuned Qwen/Qwen3.5-4B trained for math reasoning under a 8,192-token generation budget. Released as part of an anonymous ICLR 2027 submission.
Run ID (petname): awareshrimp · checkpoint step 300
Training
| Base model | Qwen/Qwen3.5-4B |
| Algorithm | GRPO with the DAPO soft overlong penalty: reward is linearly penalized (up to -1.0) over the last 512 tokens before the budget. |
Generation budget (max_new_tokens) | 8,192 |
| Data | DeepScaleR (math), 3 epochs max |
| Batch | 32 prompts × 8 rollouts per step |
| Optimizer | Adam, cosine LR schedule, peak LR 5e-7, 10 warmup steps |
| Steps | 300 |
| Reward | binary answer correctness (\boxed{} extraction) |
| Weights dtype | BF16 |
Training prompt (user turn, rendered with the base model's chat template):
Think step-by-step to solve the following problem. Output your answer inside of \\boxed{} tags.:
{problem}
Let's think step-by-step
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "budget-internalization-iclr2027/qwen3.5-4b-8k-grpo-overlong-penalty-awareshrimp-s300"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
With vLLM: vllm serve budget-internalization-iclr2027/qwen3.5-4b-8k-grpo-overlong-penalty-awareshrimp-s300
License
Inherits the license of the base model (Qwen/Qwen3.5-4B).