qwen2.5-json-extractor-p0

Qwen2.5 JSON Extractor LoRA

A LoRA adapter for Qwen2.5-1.5B that performs schema-guided structured information extraction.

Given a JSON schema and an unstructured text, the model extracts the required information and returns a valid JSON object.

The model extracts only explicitly mentioned information and never infers missing values.

---

Features

• Schema-guided extraction
• Primitive data types
• Flat JSON objects
• Missing value handling (null)
• Deterministic JSON generation
• Designed for agentic workflows and tool calling

---

Base Model

• Qwen/Qwen2.5-1.5B
• https://huggingface.co/Qwen/Qwen2.5-1.5B

---

Prompt Format

The model expects a chat conversation with the following format.

System Prompt

You are an information extraction system.

Return ONLY a valid JSON object.

Rules:
- Extract only explicitly mentioned information.
- Never infer or assume values.
- Preserve the schema exactly.
- Use null for missing values.
- Return only JSON.

User Prompt

Schema:
{
  "company": "string",
  "ceo": "string",
  "founded": "int",
  "employees": "int"
}

Text:
OpenAI was founded in 2015. Sam Altman is the CEO. The company employs approximately 4500 people.

Expected Output

{
  "company": "OpenAI",
  "ceo": "Sam Altman",
  "founded": 2015,
  "employees": 4500
}

---

Installation

Install the required packages.

pip install transformers peft accelerate torch

---

Download the Base Model

This adapter is trained on Qwen/Qwen2.5-1.5B.

The base model will be downloaded automatically the first time you run the code.

---

Load the Model

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

BASE_MODEL = "Qwen/Qwen2.5-1.5B"
LORA_MODEL = "Yellowforesty/qwen2.5-json-extractor-p0"

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

base = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    device_map="auto"
)

model = PeftModel.from_pretrained(
    base,
    LORA_MODEL
)

---

System Prompt

SYSTEM_PROMPT = """
You are an information extraction system.

Return ONLY a valid JSON object.

Rules:
- Extract only explicitly mentioned information.
- Never infer or assume values.
- Preserve the schema exactly.
- Use null for missing values.
- Return only JSON.
""".strip()

---

Prepare the Input

prompt = """
Schema:
{
    "company": "string",
    "ceo": "string",
    "founded": "int",
    "employees": "int"
}

Text:
OpenAI was founded in 2015. Sam Altman is the CEO. The company employs approximately 4500 people.
"""

---

Run Inference

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": prompt},
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=256,
        temperature=0.0,
        do_sample=False,
    )

response = tokenizer.decode(
    output[0][inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)

print(response)

---

Output

{
    "company": "OpenAI",
    "ceo": "Sam Altman",
    "founded": 2015,
    "employees": 4500
}

Inference

import os
import torch

from huggingface_hub import login
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# ==========================================
# CONFIG
# ==========================================

BASE_MODEL = "Qwen/Qwen2.5-1.5B"
LORA_MODEL = "Yellowforesty/qwen2.5-json-extractor-p0"

HF_TOKEN = os.getenv("HF_TOKEN")

if HF_TOKEN:
    login(token=HF_TOKEN)

SYSTEM_PROMPT = """
You are an information extraction system.

Return ONLY a valid JSON object.

Rules:
- Extract only explicitly mentioned information.
- Never infer or assume values.
- Preserve the schema exactly.
- Use null for missing values.
- Return only JSON.
""".strip()

PROMPT = """
Schema:
{
  "company": "string",
  "ceo": "string",
  "founded": "int",
  "employees": "int"
}

Text:
OpenAI was founded in 2015. Sam Altman is the CEO. The company employs approximately 4500 people.
""".strip()

# ==========================================
# LOAD TOKENIZER
# ==========================================

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

# ==========================================
# LOAD BASE MODEL
# ==========================================

base = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    device_map="auto",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
)

# ==========================================
# LOAD LORA
# ==========================================

model = PeftModel.from_pretrained(
    base,
    LORA_MODEL,
)

model.eval()

print("Device:", next(model.parameters()).device)

# ==========================================
# INFERENCE
# ==========================================

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": PROMPT},
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=256,
        temperature=0.0,
        do_sample=False,
    )

response = tokenizer.decode(
    output[0][inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)

print(response.strip())

---

Example

Input

Schema:
{
  "name": "string",
  "age": "int",
  "city": "string"
}

Text:
Aman, aged 31, recently moved to Chandigarh.

Output

{
  "name": "Aman",
  "age": 31,
  "city": "Chandigarh"
}

---

Benchmark

Current Version (P0)

• Primitive Types
• Flat JSON
• Null Handling

Evaluation

• JSON Validity: 100%
• Extraction Benchmark: 20 / 20

---

Limitations

Current version supports

• Primitive values
• Flat JSON objects
• Single object extraction

Not yet supported

• Nested JSON
• Arrays
• Number normalization
• Date normalization
• Multi-hop reasoning
• Planning

---

Dataset

The training dataset used to train this adapter is not included in this repository.

---

Roadmap

• ✅ P0 – Primitive JSON Extraction
• ⏳ P1 – Nested JSON
• ⏳ P2 – Number Normalization
• ⏳ P3 – Date Normalization
• ⏳ P4 – Reasoning
• ⏳ P5 – Planning

---

License

MIT