Repository role

code-tape publishes the same model family in three forms:

For browser-local inference in code-tape, use the ONNX repository. Use this repository when you need a standard Transformers checkpoint.

Intended contract

Input is a chat message containing JSON:

json
Copy code
{
  "context": {
    "fileName": "ReplayControls.tsx",
    "code": "const canSeek = durationMs > 0;",
    "runtimeOutput": "",
    "glossary": ["ReplayControls", "canSeek", "durationMs"]
  },
  "segments": [
    { "id": "subtitle-1", "startMs": 0, "endMs": 1400, "text": "这里先判断 can seek 是否可用" }
  ]
}

Expected output shape:

json
Copy code
{
  "segments": [
    { "id": "subtitle-1", "text": "这里先判断 canSeek 是否可用" }
  ],
  "chapters": [
    { "title": "判断回放是否可 seek", "startMs": 0, "endMs": 1400 }
  ]
}

Rules expected by the code-tape application:

• output JSON only, with no Markdown or explanation;
• segments contains only changed segments and may be empty;
• every returned segment id must exist in the input and must not be duplicated;
• chapter times must be monotonic, non-overlapping, and inside the subtitle timeline;
• invalid output is discarded by the application.

Usage with Transformers

python
Copy code
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "ceilf6/code-tape-subtitle-postprocessor-merged"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

messages = [
    {
        "role": "system",
        "content": (
            "You are the code-tape subtitle post-processing model.\n"
            "Only output one JSON object.\n"
            "Goal: correct ASR subtitle text for frontend/code terms and create playback chapter jump points."
        ),
    },
    {"role": "user", "content": "{\"context\":{},\"segments\":[]}"},
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=384, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training and conversion

The model was created from the code-tape subtitle post-processing LoRA workflow:

1. prepare seed records with ASR-like subtitles, code context, runtime output, and glossary terms;
2. distill strict JSON correction/chapter examples;
3. fine-tune a LoRA adapter on HuggingFaceTB/SmolLM2-135M-Instruct;
4. merge the adapter into a full model;
5. export the merged model to ONNX for browser use.

The merged checkpoint is mainly an intermediate artifact for reproducibility and export.

Evaluation

code-tape evaluates this model family with project-specific checks instead of broad language-model benchmarks:

• valid JSON object output;
• valid sparse segment references;
• glossary preservation after sparse corrections are applied back to the source subtitles;
• non-empty, ordered, non-overlapping chapter supervision for training/evaluation records;
• chapter bounds inside the subtitle timeline.

The model output must always be validated by the caller.

Limitations

• Narrowly trained for code-tape subtitle correction and chapter generation.
• Not suitable as a general chat assistant or general summarizer.
• Not an ASR model and cannot process audio directly.
• Small local models may produce malformed JSON; callers must keep a fallback path.

Privacy and security

The intended production path is the ONNX export running in the browser with @huggingface/transformers. Public browser loading does not require a Hugging Face token.

Do not put secrets, credentials, private code, or access tokens in prompts unless your inference environment is trusted.

License

Apache-2.0, following the base model license.