ankit-ml11/automerge_model

🔀 automergeAI

Automatic diff3 merge conflict resolver for Java, JavaScript, TypeScript, and C#.

Fine-tuned from flan-t5-large using LoRA on 1,200 real-world stratified merge conflicts. Given a diff3-format conflict block, the model outputs a resolved version that intelligently combines changes from both branches.

---

Model Details

Property	Value
Base model	google/flan-t5-large (783M params)
Fine-tuning method	LoRA (r=16, α=32, targets: q, v, k)
Trainable params	~2.7M (1.18% of total)
Training samples	1,200 stratified (300 per language)
Max sequence length	256 tokens (input + target)
Languages	Java · JavaScript · TypeScript · C#
Task	Text2Text Generation (diff3 conflict → resolved code)

Conflict type distribution

Type	Proportion	Description
Rich (multi-hunk true merge)	50%	Resolution differs from both branches, 2+ hunks
Complex (single-hunk merge)	30%	Resolution requires combining both sides
Simple (branch selection)	20%	One branch is clearly correct

---

Quickstart

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_id = "ankit-ml11/automerge_model"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model     = AutoModelForSeq2SeqLM.from_pretrained(model_id)
model.eval()

conflict = """<<<<<<< ours
async function getUser(id) {
    const res = await axios.get(`/api/user/${id}`);
    return res.data;
}
||||||| base
function getUser(id) { return fetch('/api/user/' + id); }
=======
async function getUser(id) {
    const res = await axios.get(`/api/user/${id}`);
    return { data: res.data, status: res.status };
}
>>>>>>> theirs"""

prompt = f"<lang_javascript> fix diff3 merge conflict by combining ALL changes from both ours and theirs:\n{conflict}"

inputs = tokenizer(prompt, return_tensors="pt", max_length=256, truncation=True)

output_ids = model.generate(
    **inputs,
    max_new_tokens=256,
    num_beams=4,
    repetition_penalty=1.5,
    no_repeat_ngram_size=4,
    length_penalty=1.2,
    early_stopping=True,
)

resolved = tokenizer.decode(output_ids[0], skip_special_tokens=True)
print(resolved)

Expected output:

async function getUser(id) {
    const res = await axios.get(`/api/user/${id}`);
    return { data: res.data, status: res.status };
}

---

Language Tokens

Prepend the appropriate language token to every prompt:

Language	Token
Java	<lang_java>
JavaScript	<lang_javascript>
TypeScript	<lang_typescript>
C#	<lang_csharp>

Prompt format:

<lang_{language}> fix diff3 merge conflict by combining ALL changes from both ours and theirs:
{conflict_block}

The conflict block must use diff3 format — it must contain all four markers: <<<<<<<, |||||||, =======, >>>>>>>

---

More Examples

Java — stream with null filter

conflict = """<<<<<<< ours
public List<String> getNames(List<User> users) {
    return users.stream().map(User::getName).collect(Collectors.toList());
}
||||||| base
public List<String> getNames(List<User> users) {
    List<String> r = new ArrayList<>();
    for (User u: users) r.add(u.getName());
    return r;
}
=======
public List<String> getNames(List<User> users) {
    return users.stream().map(User::getName).filter(Objects::nonNull).collect(Collectors.toList());
}
>>>>>>> theirs"""

prompt = f"<lang_java> fix diff3 merge conflict by combining ALL changes from both ours and theirs:\n{conflict}"

TypeScript — interface extension

conflict = """<<<<<<< ours
interface Repo<T> { findById(id: number): Promise<T>; }
||||||| base
interface Repo<T> { find(id: number): T; }
=======
interface Repo<T> {
    findById(id: number): Promise<T>;
    findAll(filter?: Partial<T>): Promise<T[]>;
    save(entity: T): Promise<T>;
}
>>>>>>> theirs"""

prompt = f"<lang_typescript> fix diff3 merge conflict by combining ALL changes from both ours and theirs:\n{conflict}"

C# — async service with caching

conflict = """<<<<<<< ours
public async Task<User> GetUserAsync(int id) {
    var cached = await _cache.GetAsync<User>($"user:{id}");
    if (cached != null) return cached;
    return await _db.Users.FindAsync(id);
}
||||||| base
public async Task<User> GetUserAsync(int id) {
    return await _db.Users.FindAsync(id);
}
=======
public async Task<User> GetUserAsync(int id) {
    var cached = await _cache.GetAsync<User>($"user:{id}");
    if (cached != null) return cached;
    var user = await _db.Users.FindAsync(id);
    await _cache.SetAsync($"user:{id}", user, TimeSpan.FromMinutes(10));
    return user;
}
>>>>>>> theirs"""

prompt = f"<lang_csharp> fix diff3 merge conflict by combining ALL changes from both ours and theirs:\n{conflict}"

---

GPU Inference (faster)

import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_id = "ankit-ml11/automerge_model"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id, torch_dtype=torch.float16)
model = model.cuda().eval()

# ... same prompt/generate code as above ...

Note: torch_dtype=torch.float16 is safe for inference only. Do not use it when fine-tuning further — AMP training requires float32 base weights.

---

Evaluation Metrics

Evaluated on a held-out set of 120 samples (10% stratified across all 4 languages).

Metric	Description	Weight in Combined Score
Exact Match	Prediction matches reference character-for-character	20%
CodeBLEU	n-gram + keyword + dataflow + line-level F1	35%
Build Rate	No conflict markers, balanced braces, no repetition loops	25%
Work Saved	Line-level edit distance vs. reference resolution	20%

---

Training curriculum

This model is the consolidation round of a multi-stage fine-tuning pipeline:

1. Per-language rounds (Java → C# → JavaScript → TypeScript), each trained on 3000 domain-specific samples
2. All-language consolidation (this model) — 10,000 balanced samples across all 4 languages to prevent catastrophic forgetting of earlier languages

---

Limitations

• Max conflict size: Conflicts longer than ~256 tokens will be truncated. Very large multi-hunk conflicts may produce incomplete output.
• Diff3 format required: Standard 2-way conflicts (<<<<<<< / ======= / >>>>>>>) without the ||||||| base section are not supported — the model needs the base section to reason about intent.
• Ambiguous conflicts: When both branches make incompatible structural changes, the model may produce plausible-looking but incorrect merges. Always review the output.
• Language coverage: Only the 4 trained languages are supported. Other languages will likely produce cross-language syntax artifacts.
• Not a replacement for human review: Treat outputs as a strong first draft, not a guaranteed correct resolution.

---

Citation

@misc{automergeai2025,
  author       = {Ankit},{Aeron}, {Bikrant}, {Bishwash}
  title        = {automergeAI: Automatic diff3 Merge Conflict Resolution with LoRA-tuned T5},
  year         = {2025},
  publisher    = {HuggingFace},
  howpublished = {\url{https://huggingface.co/ankit-ml11/automerge_model}},
}

---

License

Apache 2.0 — see LICENSE for details.