Gemma-4-E2B-it · 4k-token budget · grpo-reasoninggym · fullcoyote
RL-finetuned google/gemma-4-E2B-it trained with GRPO on Reasoning Gym tasks under a 4,096-token generation budget. Released as part of an anonymous ICLR 2027 submission.
Run ID (petname): fullcoyote · checkpoint step 300
Training
| Base model | google/gemma-4-E2B-it |
| Algorithm | GRPO (leave-one-out baseline, group reward normalization, token-level loss). Responses that hit the budget receive zero reward. |
Generation budget (max_new_tokens) | 4,096 |
| Data | Reasoning Gym tasks (procedurally generated reasoning problems, \boxed{} answers), 3 epochs max |
| Batch | 32 prompts × 8 rollouts per step |
| Optimizer | Adam, cosine LR schedule, peak LR 3e-6, 10 warmup steps |
| Steps | 300 |
| Reward | Reasoning Gym task verifier on the \boxed{} answer |
| Weights dtype | F32 |
Prompts are the Reasoning Gym task prompts, rendered with the base model's chat template.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "budget-internalization-iclr2027/gemma4-e2b-4k-grpo-reasoninggym-fullcoyote-s300"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
With vLLM: vllm serve budget-internalization-iclr2027/gemma4-e2b-4k-grpo-reasoninggym-fullcoyote-s300
License
Inherits the license of the base model (google/gemma-4-E2B-it).