TAYI

TAYI is a consumer-GPU friendly research adapter for repo-level coding-agent experiments. It is designed as a PEFT/LoRA adapter on top of Qwen/Qwen2.5-Coder-7B-Instruct, paired with the TAYI repository-level retrieval, patching, sandboxed testing, and evaluation loop.

This repository contains the TAYI LoRA adapter only. It does not include the base Qwen model weights and does not include the larger open-source teacher model used for repair experiments.

Intended Use

TAYI is intended for research on local coding assistants that operate on real repositories:

• repository context selection;
• structured patch generation;
• sandboxed test feedback;
• verified repair distillation;
• consumer-GPU adapter iteration under a 16GB VRAM ceiling.

It is not a finished frontier coding model and should not be presented as equivalent to proprietary coding agents. The current value is the adapter plus the surrounding evaluation and repair system.

Base Model

• Base model: Qwen/Qwen2.5-Coder-7B-Instruct
• Adapter type: PEFT LoRA
• LoRA rank: 16
• LoRA alpha: 32
• LoRA dropout: 0.05
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training profile: conservative QLoRA-style SFT smoke run
• Max sequence length in run manifest: 8192

Local Status

Current TAYI project readout from the local research repo:

System	Eval slice	Result
Base 7B HF student	local gold-20	13/20, pass@1 0.65
Early TAYI LoRA v0	local gold-20	14/20, pass@1 0.70
Runtime v4.4	target3	2/3, pass@1 0.6667
v5 system adapter	target3	2/3, pass@1 0.6667
v5.1 semantic adapter	target3	2/3, pass@1 0.6667, not promoted

Important caveat: this adapter is a research checkpoint. The main bottleneck is verified semantic-repair data scale, not raw GPU memory. Treat this repo as a reproducible checkpoint for continued TAYI research rather than a production release.

Loading Example

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "Qwen/Qwen2.5-Coder-7B-Instruct"
adapter = "PummyLee/TAYI"

tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter)
model.eval()

messages = [
    {"role": "system", "content": "You are TAYI, a repo-level coding assistant."},
    {"role": "user", "content": "Write a small Rust function that clamps an integer between min and max."},
]
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=256, temperature=0.2, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Project Repository

Source code, evaluation fixtures, scripts, and the development handoff brain are maintained at:

• GitHub: https://github.com/Pumnn1ayLee/TAYI
• Project state: docs/PROJECT_STATE.md
• Development brain: docs/DEVELOPMENT_BRAIN.md

Files

• adapter_model.safetensors: TAYI LoRA adapter weights.
• adapter_config.json: PEFT adapter configuration with base model set to Qwen/Qwen2.5-Coder-7B-Instruct.
• tokenizer.json, tokenizer_config.json, chat_template.jinja: tokenizer/chat-template files copied from the local training checkpoint.
• run_manifest.json: training and runtime policy manifest.
• training_args.bin: trainer metadata.
• docs/: compact handoff state from the source project.

License

This adapter repository is published under Apache-2.0 for research use. The base model is not included here; users must follow the license and terms of Qwen/Qwen2.5-Coder-7B-Instruct.