A fine-tuned Qwen2.5-Coder-1.5B-Instruct for merging code edit snippets into source files. Given an original code chunk (~35 lines) and a compact edit snippet with context markers, the model produces the merged result.

This model is designed to be used with the FastEdit toolkit, which handles AST scoping, deterministic edits, and post-processing. Using the model directly requires the exact prompt format described below.

Model variants

All variants are in this repo under subfolders:

Subfolder	Format	Size	Use case
`bf16/`	BF16 safetensors	3.2 GB	Fine-tuning, reference, GPU serving via vLLM/TGI
`mlx-8bit/`	MLX 8-bit	1.7 GB	Apple Silicon (recommended for local use)
`gguf/`	GGUF Q8_0	1.7 GB	llama.cpp, LM Studio, Ollama

Prompt format

The model expects a specific 2-message chat format. Using a different prompt will produce poor results.

System message

You are a coding assistant that helps merge code updates, ensuring every modification is fully integrated. /no_think

The /no_think suffix disables Qwen's thinking mode — without it, the model may emit thousands of reasoning tokens before producing output.

User message

Merge all changes from the <update> snippet into the <code> below.
- Preserve the code's structure, order, comments, and indentation exactly.
- Output only the updated code, enclosed within <updated-code> and </updated-code> tags.
- Do not include any additional text, explanations, placeholders, ellipses, or code fences.

<code>{original_code}</code>

<update>{update_snippet}</update>

Provide the complete updated code.

Expected output

The model outputs the merged code wrapped in <updated-code> tags:

<updated-code>
def process(data):
    try:
        result = transform(data)
        return result
    except Error as e:
        return {"error": str(e)}
</updated-code>

Complete example

Original code (what tree-sitter extracts for the target function):

python

def process(data):
    result = transform(data)
    return result

Edit snippet (what the user/agent writes):

python

def process(data):
    try:
        # ... existing code ...
    except Error as e:
        return {"error": str(e)}

Model output:

python

<updated-code>
def process(data):
    try:
        result = transform(data)
        return result
    except Error as e:
        return {"error": str(e)}
</updated-code>

The model understands # ... existing code ... markers (and language-specific variants like // ... existing code ...) as instructions to preserve the original lines in that region.

How it fits into FastEdit

In production, the model is the fallback — not the primary path:

AST scoping — tree-sitter finds the target function by name (~35 lines), so the model never sees the whole file
Deterministic text-match �� 74% of edits are resolved by matching context lines and splicing in new lines (0 tokens, <1ms)
Model merge — the remaining 26% of edits (structural changes like wrapping in try/catch, full rewrites) go to this model

The model only ever processes ~35-line chunks. It was trained on function-scoped edits, not whole files. Feeding it large inputs will degrade quality.

Using without FastEdit

If you want to use the model directly (without the toolkit), you need to:

Scope the input yourself — extract only the target function/class, not the whole file
Use the exact prompt format above — different prompts will produce different (worse) results
Parse the output — extract text between <updated-code> and </updated-code> tags
Handle edge cases — the model may emit <think> blocks (strip them), use variant tag names (<update-code>, <updated_code>), or truncate output on long functions

python

from transformers import AutoModelForCausalLM, AutoTokenizer

# BF16 (GPU / fine-tuning)
model = AutoModelForCausalLM.from_pretrained("continuous-lab/FastEdit", subfolder="bf16", torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained("continuous-lab/FastEdit", subfolder="bf16")

messages = [
    {"role": "system", "content": "You are a coding assistant that helps merge code updates, ensuring every modification is fully integrated. /no_think"},
    {"role": "user", "content": """Merge all changes from the <update> snippet into the <code> below.
- Preserve the code's structure, order, comments, and indentation exactly.
- Output only the updated code, enclosed within <updated-code> and </updated-code> tags.
- Do not include any additional text, explanations, placeholders, ellipses, or code fences.

<code>def process(data):
    result = transform(data)
    return result</code>

<update>def process(data):
    try:
        # ... existing code ...
    except Error as e:
        return {"error": str(e)}</update>

Provide the complete updated code."""}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0)
result = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
# Parse: extract text between <updated-code> and </updated-code>

Training

Base model: Qwen2.5-Coder-1.5B-Instruct
Task: Code edit merging across 13 languages

Evaluation

Tested on 22 structurally distinct edit patterns (73 cases) across 13 languages:

Path	Accuracy	Avg tokens	Avg latency
Deterministic (74% of edits)	100%	0	<1ms
Model (26% of edits)	92%	~40	~500ms
Combined	~98%	~10	~130ms

Per-language model accuracy (156-example benchmark):

Language	Accuracy
Python, Java, Kotlin, C, PHP	92%
JavaScript, TypeScript, Rust, Swift	85%
Go, C++, Ruby	77%

Limitations

Performance degrades on inputs longer than ~100 lines.
Does not handle whole-file edits well — use the FastEdit toolkit's AST scoping.
The edit snippet must use # ... existing code ... markers (or language-equivalent) for context preservation. Without markers, the model treats the entire snippet as a replacement.
Languages not in the training set may work but are untested.

License

MIT