Instructions to use juspay/xor-lite with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use juspay/xor-lite with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="juspay/xor-lite")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("juspay/xor-lite")
model = AutoModelForCausalLM.from_pretrained("juspay/xor-lite", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

PEFT
How to use juspay/xor-lite with PEFT:
```
Task type is invalid.
```
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use juspay/xor-lite with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "juspay/xor-lite"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "juspay/xor-lite",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/juspay/xor-lite

SGLang

How to use juspay/xor-lite with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "juspay/xor-lite" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "juspay/xor-lite",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "juspay/xor-lite" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "juspay/xor-lite",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use juspay/xor-lite with Docker Model Runner:
```
docker model run hf.co/juspay/xor-lite
```

XOR Lite v1.1

XOR Lite (xor-lite) is a post-trained version of openbmb/MiniCPM5-2B for typed decision tasks. It is served through a TypeSafe-compatible /v1/systemone API

Version 1.1 replaces the token readout of 1.0 with a pointer readout head and ships the LoRA adapter separately from the base weights. The adapter must be applied at runtime; the included serving bundle does this

Release

Field	Value
Version	1.1, revision tag `v1.1`
Base checkpoint	`openbmb/MiniCPM5-2B`
Base revision	`12a3808a956f869c767195e9266b59c4d21d92e2`
Architecture	Dense `LlamaForCausalLM` torso, LoRA adapter, pointer readout head
Parameters	Approximately 2.52 billion in the base, 50.2 million in the adapter, 1.1 million in the head
Released precision	BF16 base; FP32 adapter and head
License	Apache License 2.0
Packaging	Base weights plus a separate adapter and head. The adapter is applied at runtime and is not merged

Pin the published revision for reproducible use

hf download juspay/xor-lite --revision v1.1 --local-dir xor-lite

Lineage

openbmb/MiniCPM5-2B @ 12a3808
 ├─ XOR Lite 1.0 weights                ×0.3 ┐
 └─ a second MiniCPM5-2B fine-tune      ×0.7 ┘─ linear weight average (bf16) → 1.1 base
                                                 + LoRA r32, alpha 64 (applied at runtime)
                                                 + pointer readout head

The 1.1 base is a linear average of two fine-tunes of openbmb/MiniCPM5-2B: the released XOR Lite 1.0 weights (0.3) and a second fine-tune from the same campaign (0.7). The adapter and head were then trained on that base

Repository layout

Path	Contents
`config.json`, `model.safetensors`, tokenizer files	The 1.1 base model only
`decider/`	The 1.1 release checkpoint: `lora/` adapter, `head.safetensors`, tokenizer, and `strands_decider_config.json`
`serving/`	The serving bundle
`eval/decision-index-0.3.json`	The Decision Index result reported below

Loading the repository root with AutoModelForCausalLM gives the base model, not XOR Lite 1.1. Do not merge the adapter into the base: the bf16 base weights cannot represent part of the FP32 adapter delta, and a merged model gives different answers

Interface

The server accepts a state and a map of typed questions

noul: binary probability
choice: categorical decision, full probability distribution, and confidence
score: expected ordinal score, full probability distribution, legend, and confidence

The readout head scores each option from the model's hidden state at the end of that option. choice accepts 2 to 255 options and score accepts 2 to 10 levels; larger requests are rejected with HTTP 422. The state and questions are text or JSON; image and video inputs are not supported

Quick start

hf download juspay/xor-lite --revision v1.1 --local-dir xor-lite

(cd xor-lite/serving && sha256sum -c xor-lite-1.1-serving.tar.gz.sha256)

mkdir -p xor-lite-runtime
tar -xzf xor-lite/serving/xor-lite-1.1-serving.tar.gz -C xor-lite-runtime --strip-components=1

cd xor-lite-runtime
cp .env.example .env
sed -i "s|^MODEL_DIR=.*|MODEL_DIR=$(cd ../xor-lite && pwd)|" .env
./run.sh

run.sh verifies every model file, builds the API image, and runs a smoke test. When it succeeds, the API is available at http://127.0.0.1:30002/v1/systemone

Decision Index 0.3

Score	Value
Decision Index 0.3	37.51
Raw index	53.66
Breadth skill	34.74

Area	Skill	Raw
Knowledge & Reasoning	0.1828	0.3879
Language Understanding	0.4156	0.5936
Retrieval & Classification	0.5006	0.6439
Tools & Automation	0.5664	0.6041
Arts & Human Taste	0.1670	0.4353

This is a self-run result on the public Decision Index 0.3 suite, with complete coverage. It is not an official board entry. The run scored the released weights in-process on one GPU. Its noul answers were then recomputed with the released serving bundle, because the run's own noul readout reported the larger of the two probabilities instead of the probability of true. That correction changed one benchmark and raised the index from 37.05 to 37.51. A replay of 1,000 sampled cases through the released bundle matched the run's choice answers in 2,005 of 2,007 questions, with the other two within 0.015. The full per-benchmark result is in eval/decision-index-0.3.json

The training data includes ChessBench-style rows, and ChessBench is one of the index benchmarks. Results on that benchmark should not be read as held-out performance

Validated runtime

Setting	Value
Runtime	PyTorch 2.13.0 (CUDA 13.0), Transformers 5.14.1, PEFT at commit `194cb8f`
Validated GPU	1 x NVIDIA RTX PRO 6000 Blackwell Server Edition, 96 GB
NVIDIA driver	`595.91.07`
Workers	1
Shared-prefix cache	Off
Probability temperature	`1.0` for `noul`, `choice`, and `score`

The loaded service used about 21.9 GB of GPU memory after serving the replay above. Latency has not been measured for this release

The service binds to localhost by default and does not require an API key. Remote deployments must add authentication, TLS, rate limits, and request-size limits at the ingress layer

Changes from 1.0

The base is now the linear weight average described above; 1.0 contributes 0.3 of it
Answers come from a pointer readout head instead of single-token log probabilities, and the temperature is 1.0 instead of 100.0
The adapter and head ship as separate files and are applied at runtime. Deployments that load the repository root directly in SGLang or Transformers get the base model only
Responses include confidence for choice and score, and legend for score
The serving bundle runs the model in PyTorch rather than SGLang

Known provenance gaps

The second fine-tune in the base average was identified from its weights: recomputing the 0.7/0.3 average reproduces every base tensor bit for bit. Its training configuration is not part of this release
The adapter was saved with a development build of PEFT (0.19.2.dev0, commit 194cb8f); the serving bundle installs that exact commit

Downloads last month: 672

Safetensors

Model size

3B params

Tensor type

BF16

Model tree for juspay/xor-lite

Base model

openbmb/MiniCPM5-2B

Adapter

(30)

this model