TeachMaster Coder

Turn lesson text into animated teaching visuals.

Qwen3.8-27B · LoRA SFT · Remotion TSX

Base model Adapter Format Weights

从课程文字出发,生成包含标题、图片、流程图、公式与动画的可执行 TSX。

Gallery · Training · Quick Start · Limitations

Generated teaching scene: end-to-end speech recognition

Real model output. Gallery images are Remotion renders of this adapter's generated TSX. No manual layout repairs were applied. The speech-recognition illustrations are existing dataset assets referenced by the code; this model did not generate those images.

What It Does

TeachMaster Coder is a task-specific adapter for Chinese educational animation code generation. It takes lesson Markdown, image-planning JSON and presentation requirements, and produces complete React/Remotion TSX.

Input Output
Lesson content and terminology Structured teaching scenes
Image descriptions, ratios and asset paths Image placement with Img and staticFile
Canvas, timing and visual requirements Components, layout styles and frame-based animation

This repository contains the LoRA adapter, not the full base model. Use it with Qwen/Qwen3.8-27B. Source training data and image assets are not distributed here.

Generated Gallery

01 / From Audio to Understanding

An illustration, definition panel, processing pipeline and side-by-side comparison share one 1920 × 1080 canvas.

End-to-end speech recognition: generated scene 1

02 / Complex Environments, Structured Explanations

Image-and-text composition with three explanatory blocks, an audio-processing pipeline and application cards.

Microphone arrays and multilingual recognition: generated scene 2

03 / A Concept Made Visible

A text-only lesson becomes an application overview and a highlighted Bayesian formula. This example uses no image assets.

Prior probability: generated summary and applications

See the remaining two scenes, including the sparse layout

The first scene defines prior probability and organizes three sources of prior knowledge.

Prior probability: definition

The mechanism scene illustrates a remaining weakness: content is organized, but too much space is left unused in the middle of the frame.

Prior probability: mechanism and excessive whitespace

How These Examples Were Checked

Check Observation
Evaluation data 2 held-out pages; neither source job appears in this run's training split
Generation Greedy decoding, thinking disabled, maximum 16,384 new tokens
Completion Both outputs ended normally with EOS; 5,267 and 5,322 output tokens
Rendering All 5 selected scene-end frames rendered without runtime errors
Layout inspection No obvious clipping or text overlap in those screenshots
Scope A small qualitative smoke test, not a full benchmark or all-frame audit

The images above are selected still frames, not evidence of narration alignment or animation quality across the full timeline. The model produced 16- and 21-second compositions despite a request for approximately 40 seconds.

Training at a Glance

Configuration
Base model Qwen/Qwen3.8-27B
Transformers architecture Qwen3_5ForConditionalGeneration
Training / validation 11,450 / 515 examples, split by source job
Objective Supervised TSX generation; prompt tokens excluded from loss
Thinking Disabled; no chain-of-thought supervision
LoRA Rank 64 · alpha 128 · dropout 0.05
Target modules q_proj, k_proj, v_proj, o_proj
Precision / hardware BF16 · 8 × NVIDIA A800 80GB
Sequence cutoff 12,288 tokens
Effective batch 32 = 8 GPUs × microbatch 1 × accumulation 4
Schedule 2 epochs · 716 steps · peak LR 5e-5 · cosine decay · 3% warmup
Training runtime Approximately 12.5 hours, excluding final evaluation
Final mean training loss 0.185514
Final validation loss 0.173346

Length coverage: 479 training examples and 19 validation examples exceed the 12,288-token cutoff and are subject to truncation. The run does not cover every example at its original full length.

Learning Curves

Training loss Validation loss
Training loss Validation loss

Validation loss improved from 0.1851 at step 100 to 0.1733 at completion. Token loss measures code prediction; it does not directly measure layout aesthetics, content correctness or animation timing.

Quick Start

The tested runtime used Transformers 5.8, PEFT 0.18.1, PyTorch with BF16 support, flash-linear-attention and causal-conv1d. Install versions compatible with your CUDA/PyTorch environment. This base model uses the Qwen3.5 architecture implementation despite its Qwen3.8 model name.

import torch
from peft import PeftModel
from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration

base_id = "Qwen/Qwen3.8-27B"
adapter_id = "posprivacy/teachmaster-qwen38-27b-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base = Qwen3_5ForConditionalGeneration.from_pretrained(
    base_id,
    dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="sdpa",
)
model = PeftModel.from_pretrained(base, adapter_id).eval()

messages = [
    {
        "role": "system",
        "content": (
            "你是 Remotion TSX 代码生成器。只输出完整、可运行的 TSX。"
            "画布为 1920x1080,fps=30。使用 Sequence 组织场景,"
            "确保文字、图形和图片不出界、不重叠。"
        ),
    },
    {
        "role": "user",
        "content": '''【课程讲义 Markdown】
先验概率描述观测新数据之前对模型参数的初始判断。
它可以来自历史经验、科学理论和逻辑推演。

【图片规划 JSON】
{"needs_image": false, "images": []}

【当前分页文件名】
5_1.md

请用定义、流程图和应用示例组织画面。''',
    },
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    enable_thinking=False,
    return_tensors="pt",
    return_dict=True,
).to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=16384,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )

tsx = tokenizer.decode(
    output[0, inputs["input_ids"].shape[-1]:],
    skip_special_tokens=True,
)
print(tsx)

This is a compact loading example. The gallery used a longer reconstructed coder system prompt specifying scene structure, typography, color palette, image integration and animation conventions. The short example above is not an exact reproduction of the gallery input.

Working With Images

Supply an image-planning JSON and an explicit path convention in the user message:

【图片规划 JSON】
{"needs_image": true, "images": [{"index": 1, "ratio": "16:9", "description": "端到端语音识别结构图"}]}

【当前分页文件名】
24_1.md

图片路径使用 staticFile("assets/24_1/image_<index>.png")。

Place the actual asset at public/assets/24_1/image_1.png in your Remotion project. The adapter generates the reference and layout; image generation and asset delivery are separate stages.

What Still Needs Work

  • Space utilization: oversized cards, small text and excessive whitespace remain in some scenes.
  • Timing compliance: requested composition duration is not reliably followed.
  • Visual diversity: the gallery uses a shared template and palette supplied by the prompt; broader style generalization has not been established.
  • Content accuracy: generated explanations and formulas still need subject-matter review.
  • Runtime reliability: successful rendering of two samples does not establish a general compile-success rate.
  • Animation and narration: these screenshots do not test all frames, narration generation or semantic audio alignment.

Generated TSX should be compiled, executed and visually inspected before use. The next useful quality checks are full-timeline rendering and DOM-based measurements of clipping, overlap and occupied area.

Repository Contents

Files Purpose
adapter_model.safetensors, adapter_config.json LoRA weights and base-model reference
tokenizer*, chat_template.jinja, processor_config.json Tokenization and supporting configuration
train_results.json, eval_results.json Recorded training and evaluation metrics
training_loss.png, training_eval_loss.png Learning curves
assets/ Rendered gallery screenshots
checksums.json Adapter SHA256 checksum

The original base-model weights, training corpus and source image assets are not included. Refer to the base model's model card and license for its usage terms.

Downloads last month
5,775
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for posprivacy/teachmaster

Base model

Qwen/Qwen3.8-27B
Adapter
(373)
this model