CoolFace
Modelpublic

budget-internalization-iclr2027/qwen3.5-4b-4k-sft-pumpedjackal-s90

sourceHugging Faceapache-2.0updated 2d agoView on Hugging Face
0likes12downloads
Model Card

Qwen3.5-4B · SFT on 4k RFT data · pumpedjackal

Supervised finetune of Qwen/Qwen3.5-4B on its own reward-filtered solutions to DeepScaleR math problems (rejection-sampling finetuning), restricted to responses of at most 4k tokens. Released as part of an anonymous ICLR 2027 submission.

Run ID (petname): `pumpedjackal` · checkpoint step 90

Training

Base modelQwen/Qwen3.5-4B
AlgorithmSFT (next-token loss on the response)
Dataself-generated, correctness-filtered Qwen3.5-4B solutions to DeepScaleR problems; subset 4k (responses of at most 4k tokens), one solution per problem
Max sequence length18,432
Batch32 sequences per step
OptimizerAdam, cosine LR schedule, peak LR 5e-06
Epochs3
Steps90
Weights dtypeBF16

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

python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "budget-internalization-iclr2027/qwen3.5-4b-4k-sft-pumpedjackal-s90"
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-4k-sft-pumpedjackal-s90

License

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