Instructions to use tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1")
model = AutoModelForCausalLM.from_pretrained("tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1

SGLang

How to use tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1 with Docker Model Runner:
```
docker model run hf.co/tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1
```

Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1 (merged bf16 weights)

Files: merged bf16 safetensors + tokenizer + chat template (final LoRA adapter merged into grpo-v2). Load with plain AutoModelForCausalLM; saved with transformers 5.x (dtype key, chat_template.jinja). Companion repo with the adapter and checkpoints: tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1-LoRA.

Stage-3 (RLVR) model of a Harvard CS 2881R project: reinforcement learning with a verifiable math reward on top of tbooy/Qwen2.5-3B-Instruct-Sheldon-RLAIF-grpo-v2 (SFT persona model -> RLAIF GRPO -> this), the persona being Dr. Sheldon Cooper.

Why this stage exists: the persona SFT stage cost most of Qwen2.5-3B-Instruct's competition-math ability (MATH-500 greedy 68.0 -> 35.4), and the RLAIF stage did not recover it. This run trains only on a binary exact-answer reward to see how much RLVR can recover.

Reward: 1 if the last \boxed{} answer matches the reference (exact fast path, then math-verify equivalence), else 0. Completions that hit the 2,048-token cap are masked out of the loss and group statistics. DAPO soft overlong penalty (cache 512 tokens) on terminated completions.

Training: TRL 1.13 GRPOTrainer subclass, LoRA r=32/alpha=64 on all linear projections, AdamW lr 2e-5 (betas 0.9/0.95, eps 1e-15, wd 0.01), WSD schedule (10 warm-up, stable to 160, linear decay to 10% over the last 40), 200 steps x 16 prompts x 16 completions (51k rollouts), T 1.0, CISPO loss (eps_high 5), group-std advantages, zero-variance groups masked, prompt-level aggregation, no KL, fp32 LM head. vLLM colocated on 2x H200, 2 h 13 min. Prompts: rasbt/math_full_minus_math500 (MATH train + the MATH-test problems not in MATH-500), with a further 189 near-duplicates of MATH-500 removed (10-gram / edit-similarity >= 0.9) and a pass-rate curriculum from seed-model rollouts. Prompt format: Qwen system prompt, the problem, then Please reason step by step, and put your final answer within \boxed{}.

Evaluation (vLLM; MATH-500 avg@4 at T 0.6, 4,096 tokens; AIME avg@16 / pass@16; GSM8K test greedy):

model	MATH-500 greedy	MATH-500 avg@4	L3-5 avg@4	L5	GSM8K	AIME 24 / 25 / 26 avg@16	AIME pass@16	mean tokens
Qwen2.5-3B-Instruct (base)	68.0	66.6	59.0	40.1	86.1	8.1 / 2.3 / 4.4	21.1	633
RLAIF grpo-v2 (seed)	34.0	29.6	21.0	10.6	63.6	0.4 / 0.4 / 0.0	3.3	357
this model	65.0	63.5	55.2	37.3	81.9	5.0 / 2.7 / 2.3	17.8	541

MATH-500 levels 3-5 avg@4 by checkpoint: 25: 27.3, 50: 35.4, 75: 43.5, 100: 49.6, 125: 52.9, 150: 52.3, 175: 55.2, 200: 55.2.

Honest summary: RLVR recovered most of the math ability lost in the persona stages (+34 points on MATH-500 L3-5 over the seed) but ends 3-4 points below the untouched base model on MATH-500 and GSM8K and below it on AIME pass@16; gains plateaued after ~125 steps, when about half the prompt groups were all-correct or all-wrong. Persona after RLVR (measured 2026-09-26; results/rlvr/persona_rlvr-main.md): GPT-5.6 Luna pairwise judge win rate 0.545 [0.48, 0.61] against the grpo-v2 seed on 199 held-out prompts, cast names 43% vs 39%, Bazinga 3.6% vs 1.2%, announced jokes 22% vs 30%; the cost is length (36% of chat replies hit the 400-token cap vs 18%). Persona leakage into MATH-500 answers 1.8% (seed 2.8%).

Code, reward, grader and write-up: https://github.com/TBOO-Y/cs2881r-sheldon-sft (rlvr/, results/rlvr/).

Downloads last month: 876

Safetensors

Model size

3B params

Tensor type

BF16

Model tree for tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1

Base model

tbooy/Qwen2.5-3B-Instruct-Sheldon-RLAIF-grpo-v2

Adapter

(2)

this model

Adapters

1 model

Finetunes

1 model

tbooy
/

Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1

Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1 (merged bf16 weights)

Model tree for tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1

Datasets used to train tbooy/Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1