Training Data

Dataset summary:

1,036 instruction/code examples.
124 unique K-12 science topics.
Synthetic dataset generated via parallel agent workflows and validated into JSONL format.

Each example contains:

instruction
topic
grade_level
p5js_code (full runnable sketch)

Data style constraints emphasized:

600x400 canvas
setup() and draw() structure
labels/annotations for explanation
interactive and animated behavior

Training Procedure

Hardware and runtime:

4x NVIDIA A100-SXM4-80GB
~171.9 seconds total training runtime
multi-GPU training via accelerate

LoRA config:

r=64
lora_alpha=128
lora_dropout=0.05
target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Optimization config:

epochs: 3
per-device batch size: 4
gradient accumulation: 2
effective batch size: 32 (4 GPUs)
learning rate: 2e-4
scheduler: cosine, warmup ratio 0.05
weight decay: 0.01
max sequence length: 2048
precision: bf16
optimizer: AdamW

Training/Eval Snapshot

Train loss (logged): 0.9090 -> 0.4950 (step 10 to step 90)
Train run average loss (train_metrics.json): 0.5917
Eval loss (step 50): 0.6164
Reported token accuracy during run: up to ~85.6%

These are training-time indicators, not full benchmark performance across external test sets.

Quick Start

Transformers + PEFT

python

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE = "Qwen/Qwen3-0.6B"
ADAPTER = "mr-dee/qwen3-p5js-physics-lora"

tokenizer = AutoTokenizer.from_pretrained(BASE, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    BASE,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, ADAPTER)

messages = [
    {
        "role": "system",
        "content": (
            "You are a p5.js animation expert for K-12 physics education. "
            "Return complete, runnable JavaScript."
        ),
    },
    {
        "role": "user",
        "content": "Create an interactive p5.js sketch that teaches projectile motion.",
    },
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    temperature=0.4,
    top_p=0.9,
)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

vLLM Serving (LoRA)

bash

vllm serve Qwen/Qwen3-0.6B \
  --enable-lora \
  --lora-modules p5js=mr-dee/qwen3-p5js-physics-lora \
  --tensor-parallel-size 2 \
  --max-model-len 2048

Then query:

bash

curl http://127.0.0.1:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "p5js",
    "messages": [
      {"role":"system","content":"You are a p5.js animation expert for K-12 physics education."},
      {"role":"user","content":"Create an animation showing wave interference with labeled nodes and antinodes."}
    ],
    "max_tokens": 800,
    "temperature": 0.4
  }'

Important:

keep prompt_tokens + max_tokens <= max_model_len
otherwise vLLM returns HTTP 400 validation errors

Limitations and Risks

Output is not guaranteed bug-free JavaScript; review before use.
Physical explanations may be simplified or occasionally incorrect.
Performance drops for domains outside training distribution.
Small base model can struggle with very long or highly complex simulations.

Responsible Use

Use with human review in educational settings.
Validate scientific correctness before presenting to students.
Sandbox or lint generated code before running in production applications.

Framework Versions

PEFT: 0.18.1
Transformers: 4.57.6
TRL: 0.27.2
PyTorch: 2.9.1
Accelerate: 1.12+

Citation

If you use this adapter, cite:

bibtex

@misc{qwen3_p5js_physics_lora_2026,
  title  = {qwen3-p5js-physics-lora},
  author = {mr-dee},
  year   = {2026},
  url    = {https://huggingface.co/mr-dee/qwen3-p5js-physics-lora}
}