CoolFace
Modelpublic

budget-internalization-iclr2027/qwen3.5-4b-16k-grpo-reasoninggym-restedmarlin-s300

sourceHugging Faceapache-2.0updated 2d agoView on Hugging Face
0likes17downloads
Model Card

Qwen3.5-4B · 16k-token budget · grpo-reasoninggym · restedmarlin

RL-finetuned Qwen/Qwen3.5-4B trained with GRPO on Reasoning Gym tasks under a 16,384-token generation budget. Released as part of an anonymous ICLR 2027 submission.

Run ID (petname): `restedmarlin` · checkpoint step 300

Training

Base modelQwen/Qwen3.5-4B
AlgorithmGRPO (leave-one-out baseline, group reward normalization, token-level loss). Responses that hit the budget receive zero reward.
Generation budget (max_new_tokens)16,384
DataReasoning Gym tasks (procedurally generated reasoning problems, \boxed{} answers), 3 epochs max
Batch32 prompts × 8 rollouts per step
OptimizerAdam, cosine LR schedule, peak LR 5e-7, 10 warmup steps
Steps300
RewardReasoning Gym task verifier on the \boxed{} answer
Weights dtypeBF16

Prompts are the Reasoning Gym task prompts, rendered with the base model's chat template.

Usage

python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "budget-internalization-iclr2027/qwen3.5-4b-16k-grpo-reasoninggym-restedmarlin-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-16k-grpo-reasoninggym-restedmarlin-s300

License

Inherits the license of the base model (Qwen/Qwen3.5-4B).