GRPO checkpoint of Qwen/Qwen3.5-4B, trained on instruction-following (IFEval-style) prompts with a verifiable reward. This repo is the export of global step 200 of the run Qwen3.5-4B-GRPO-01R-2048-ifeval-v0-think.

Training

Trained with verl (0.9.0.dev) on 4 GPUs, FSDP2 actor + vLLM rollout.


Algorithm	GRPO (`adv_estimator=grpo`), 4 rollouts per prompt
Steps	200
Train batch size	128 prompts
Mini / micro batch	64 / 2 per GPU
Max prompt / response length	1024 / 2048 tokens
Learning rate	1e-6
KL	`use_kl_loss=True`, `kl_loss_coef=0.001`, `low_var_kl`; no KL in reward
Entropy coefficient	0
Rollout	vLLM, TP=1, temperature per verl defaults

Prompts are in reasoning ("think") format; responses are scored by a rule-based instruction-following checker.

Precision

The weights are stored as float32 — this is verl's raw training export, not a post-training cast. config.json declares dtype: bfloat16, so transformers casts to bf16 on load by default. Pass dtype="float32" if you want the stored precision.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "ZetaRRR/Qwen3.5-4B-VerIH-step200"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="auto", device_map="auto")

messages = [{"role": "user", "content": "Write a haiku about gradients. Use no commas."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=2048)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

Qwen3.5-4B is a hybrid-attention (GatedDeltaNet + full attention) multimodal architecture; the vision tower is carried over from the base model unchanged, as RL touched text only.