yinglong-intent-qwen3.5-2b

A LoRA adapter for Qwen/Qwen3.5-2B (Apache-2.0). It reads what a person types to their AI agent (for example "pay my rent" or "never spend more than 500 a month on food") and proposes a JSON document: either an intent (what to do) or a mandate (the spending rules). It is the intent understanding step of YINGLONG, an AI agent wallet that pays only inside rules the owner locked.

Important: the output is only a proposal. The model never writes an address or moves money. The YINGLONG service validates the JSON against a schema, grounds every field in the owner's own data, and the owner seals the result; the on-chain contracts then enforce it.

Base model note. adapter_config.json names Qwen/Qwen3.5-2B as the model the adapter was trained on. AutoPeftModelForCausalLM will download that base automatically.

How to use

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

repo = "ABRUBAB/yinglong-intent-qwen3.5-2b"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoPeftModelForCausalLM.from_pretrained(repo, device_map="auto")
msgs = [{"role": "user", "content": "Pay my rent on the first of every month, up to 900."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=256)[0][ids.shape[1]:], skip_special_tokens=True))

Use the same system prompt and JSON schema as the YINGLONG service (ml/intent and packages/mandate in the project repository) to get the exact output format. Run it in non-thinking mode.

Results (held-out test split, 1255 items)

Valid JSON 99.9%, exact match 88.7% (86.8% to 90.3%), slot F1 97.9%.

Slice	Items	Valid JSON	Exact match	Slot F1
overall	1255	99.9%	88.7%	97.9%
task=intent	917	100.0%	86.7%	97.2%
task=mandate	338	99.7%	94.1%	98.9%
language=bn	318	100.0%	80.8%	96.1%
language=en	468	100.0%	95.1%	99.0%
language=mixed	283	99.7%	82.3%	97.0%
language=zh	186	100.0%	95.7%	99.5%

Training

LoRA rank 16, alpha 16, 2 epochs, learning rate 0.0002, effective batch 16, evaluation context 2048 tokens. Data: synthetic only, written by Qwen/Qwen3-14B-AWQ; the labels were sampled by code, checked against the JSON schema and verified by a round trip (the mandate compiler and decompiler). The training data is not published. Languages: English, Bengali, Chinese and code-mixed.

Limits

• Exact match is lower for code-mixed text than for English; always validate and ask the owner to confirm.
• It cannot be trusted with addresses, amounts it did not read, or anything outside the schema.
• Built for a hackathon demonstration with test units and test networks; no third-party audit.

Part of YINGLONG, by team Shadow Chaser (Bangladesh), for the 2026 Shenzhen International FinTech Competition. Code and evidence: https://github.com/ABRUBAB/YINGLONG . Live demo: https://app.yinglong-agent.tech .