nabin2004/qwen3-8b-manimator-gold-sft

Qwen3-8B Manimator Gold SFT Dataset

Gold-standard Supervised Fine-Tuning (SFT) dataset for Qwen3-8B (also compatible with Qwen 2.5 7B/14B/32B, LLaMA 3.1/3.3, and DeepSeek) designed to produce high-precision, pedagogically structured, narrated mathematical animations using Manim Community Edition (v0.19+) and Manim Voiceover.

📊 Dataset Summary

• Total Samples: 733 curated, deduplicated gold samples
• Format: OpenAI / ChatML 3-turn format (system -> user -> assistant)
• Language: English / Python (Manim CE v0.19+)
• Primary Split: train.jsonl

from datasets import load_dataset

dataset = load_dataset("nabin2004/qwen3-8b-manimator-gold-sft")
print(dataset["train"][0]["messages"])

---

🎯 Target Capabilities

This dataset trains language models to:

1. Plan Before Coding: Generate an 8-key structured <Plan> block before writing any code (goal, skills, layout, camera, dynamics, math, narration, validate).
2. Audio-Visual Parity: Perfectly synchronize speech narration with visual animation using inline <bookmark mark='...'/> tags and self.wait_until_bookmark("...").
3. Frame-Relative Safety: Enforce fit_in_frame(mob, w_frac=0.9, h_frac=0.9) and MARGIN = 0.5 to eliminate off-screen clipping and visual overlap.
4. Clean Zero-Shot Prompting:
   • Every sample uses concise, natural user instructions without in-context skill bloat, training the model to inherently know Manim CE and voiceover patterns.

---

📂 Data Sources & Curation

Every sample in this dataset was vetted to guarantee zero hallucinated APIs and exact contract compliance:

Source	Samples	Format	Description
Local Canonical Validated	149	Clean Zero-Shot Prompts	Human-reviewed, linter-validated gold canonicals (data/canonical/validated.jsonl)
Qwen Chat Exports	279	Clean Zero-Shot Prompts	Fresh multi-batch generations aligned from Qwen chat sessions
nabin2004/qwen-Manimator-1-sft-data	305	Clean Zero-Shot Prompts	Core curated SFT examples with verified <Plan> and bookmark synchronization

Quality Filter: Noisy legacy samples lacking <Plan> blocks, duration-multiplier audio timing, or unresolvable internal imports (e.g. AOS-Narrated-Manim-400) were completely eliminated.

---

🏗️ Message Schema

Each line in train.jsonl has the standard format:

{
  "messages": [
    {
      "role": "system",
      "content": "You are an expert Python programmer and mathematics educator specializing in ManimCE and Manim Voiceover. You create high-quality, pedagogically rich, narrated animations. You always use `VoiceoverScene`, `GTTSService`, and precise bookmark timing (`<bookmark>` tags and `wait_until_bookmark`) to sync audio with visual elements. Wrap your code in ```python ... ``` blocks."
    },
    {
      "role": "user",
      "content": "Create an interactive Manim CE animation demonstrating: Cycloid Generation with TracedPath and rolling circle."
    },
    {
      "role": "assistant",
      "content": "<Plan>\ngoal: ...\nskills: ...\nlayout: ...\ncamera: ...\ndynamics: ...\nmath: ...\nnarration: ... <bookmark mark='intro'/> ...\nvalidate: ...\n</Plan>\n\n```python\nfrom manim import *\n...\n```"
    }
  ]
}

---

🚀 Quickstart: Fine-Tuning Qwen3-8B

Using Hugging Face TRL (SFTTrainer):

import torch
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer
from trl import SFTTrainer, SFTConfig

model_id = "Qwen/Qwen2.5-7B-Instruct"  # or Qwen3-8B upon release

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

dataset = load_dataset("nabin2004/qwen3-8b-manimator-gold-sft")

training_args = SFTConfig(
    output_dir="./qwen3-manimator-gold",
    max_seq_length=8192,
    num_train_epochs=3,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    learning_rate=1e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    bf16=True,
    logging_steps=10,
    save_strategy="epoch",
    dataset_text_field="messages",
)

trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=dataset["train"],
)

trainer.train()

Using Unsloth (Ultra-Fast 2x QLoRA):

from unsloth import FastLanguageModel
from datasets import load_dataset
from trl import SFTTrainer, SFTConfig

max_seq_length = 8192
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="Qwen/Qwen2.5-7B-Instruct",
    max_seq_length=max_seq_length,
    load_in_4bit=True,
)

model = FastLanguageModel.get_peft_model(
    model,
    r=64,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    lora_alpha=128,
    lora_dropout=0,
)

dataset = load_dataset("nabin2004/qwen3-8b-manimator-gold-sft")

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset["train"],
    dataset_text_field="messages",
    max_seq_length=max_seq_length,
    dataset_num_proc=2,
    args=SFTConfig(
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,
        warmup_steps=10,
        max_steps=250,
        learning_rate=2e-4,
        fp16=False,
        bf16=True,
        logging_steps=1,
        output_dir="outputs",
    ),
)

trainer.train()

---

📜 License

Apache 2.0. Open for commercial, educational, and research use.