rank1-14b-awq: Quantized Model for Test-Time Compute Reranking

📄 Paper | 🚀 GitHub Repository

rank1-14b-awq is a quantized version of the rank1-14b model. This AWQ-quantized 14B parameter model maintains the reasoning capabilities of the original model while requiring less memory and providing faster inference. The model is trained from the Qwen2.5-14B base model and leverages test-time compute to generate reasoning chains before deciding if a document is relevant to a query.

Model Description

rank1 introduces a novel approach to information retrieval by generating explicit reasoning chains before making relevance judgments. Unlike traditional rerankers that directly output scores, rank1:

1. Receives a query and document pair
2. Generates a reasoning chain within a <think>...</think> section
3. Makes a binary relevance judgment (true or false)
4. Returns a confidence score based on the logits of the true/false tokens

This approach helps the model break down complex relevance decisions into logical steps, improving performance across diverse retrieval tasks.

Quantization Details

This model uses Activation-aware Weight Quantization (AWQ) to reduce the model size while maintaining performance. Compared to the full-precision model, this quantized version:

• Requires less GPU memory
• Offers faster inference times
• Maintains comparable accuracy on retrieval tasks

Model Family

Model	Base	Description
rank1-7b	Qwen2.5-7B	Smaller variant (7B parameters)
rank1-14b	Qwen2.5-14B	Full-precision version (14B parameters)
rank1-32b	Qwen2.5-32B	Largest variant (32B parameters)
rank1-mistral-2501-24b	Mistral-Small 2501 24B	Trained from Mistral base
rank1-llama3-8b	Llama 3.1 8B	Trained from Llama 3.1 base

Quantized Variants

Model	Description
rank1-7b-awq	Quantized version of rank1-7b
rank1-14b-awq	Current model - Quantized version of rank1-14b
rank1-32b-awq	Quantized version of rank1-32b
rank1-mistral-2501-24b-awq	Quantized version of rank1-mistral-24b
rank1-llama3-8b-awq	Quantized version of rank1-llama3-8b

Associated Data and Resources

Resource	Description
rank1-r1-msmarco	All R1 output examples from MS MARCO
rank1-training-data	Training data used for rank1 models
rank1-run-files	Pre-computed run files for use in top 100 doc reranking
GitHub Repository	Official rank1 repository

Usage

Note that official usage is found on the Github and accounts for edge cases. But for simple use cases the minimal example below works.

Click to expand: Minimal example with vLLM

from vllm import LLM, SamplingParams
import math

# Initialize the model with vLLM
model = LLM(
    model="jhu-clsp/rank1-14b-awq",
    tensor_parallel_size=1,  # Number of GPUs
    trust_remote_code=True,
    max_model_len=16000,     # Context length
    gpu_memory_utilization=0.9,
    dtype="auto",  # Will use the appropriate quantized dtype
)

# Set up sampling parameters
sampling_params = SamplingParams(
    temperature=0,
    max_tokens=8192,
    logprobs=20,
    stop=["</think> true", "</think> false"],
    skip_special_tokens=False
)

# Prepare the prompt
def create_prompt(query, document):
    return (
        "Determine if the following passage is relevant to the query. "
        "Answer only with 'true' or 'false'.\n"
        f"Query: {query}\n"
        f"Passage: {document}\n"
        "<think>"
    )

# Example usage
query = "What are the effects of climate change?"
document = "Climate change leads to rising sea levels, extreme weather events, and disruptions to ecosystems. These effects are caused by increasing greenhouse gas concentrations in the atmosphere due to human activities."

# Generate prediction
prompt = create_prompt(query, document)
outputs = model.generate([prompt], sampling_params)

# Extract score
output = outputs[0].outputs[0]
text = output.text
final_logits = output.logprobs[-1]

# Get token IDs for "true" and "false" tokens
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("jhu-clsp/rank1-14b-awq")
true_token = tokenizer(" true", add_special_tokens=False).input_ids[0]
false_token = tokenizer(" false", add_special_tokens=False).input_ids[0]

# Calculate relevance score (probability of "true")
true_logit = final_logits[true_token].logprob
false_logit = final_logits[false_token].logprob
true_score = math.exp(true_logit)
false_score = math.exp(false_logit)
relevance_score = true_score / (true_score + false_score)

print(f"Reasoning chain: {text}")
print(f"Relevance score: {relevance_score}")

Click to expand: Usage with AutoGPTQ/AWQ

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load the tokenizer and quantized model
tokenizer = AutoTokenizer.from_pretrained("jhu-clsp/rank1-14b-awq")
model = AutoModelForCausalLM.from_pretrained(
    "jhu-clsp/rank1-14b-awq",
    device_map="auto",
    trust_remote_code=True
)

# Prepare the prompt
query = "What are the effects of climate change?"
document = "Climate change leads to rising sea levels, extreme weather events, and disruptions to ecosystems. These effects are caused by increasing greenhouse gas concentrations in the atmosphere due to human activities."

prompt = f"Determine if the following passage is relevant to the query. Answer only with 'true' or 'false'.\nQuery: {query}\nPassage: {document}\n<think>"

# Generate the reasoning chain and relevance judgment
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.0,
        return_dict_in_generate=True,
        output_scores=True,
        pad_token_id=tokenizer.eos_token_id
    )

# Process the output
generated_text = tokenizer.decode(outputs.sequences[0], skip_special_tokens=False)
reasoning_chain = generated_text.split("<think>")[1].split("</think>")[0].strip()
relevance_judgment = "true" if "true" in generated_text.split("</think>")[1].strip().lower() else "false"

print(f"Reasoning chain: {reasoning_chain}")
print(f"Relevance judgment: {relevance_judgment}")

Performance

rank1-14b-awq demonstrates strong performance on retrieval benchmarks while offering faster inference and lower memory requirements than the full-precision model. The quantization process preserves the model's ability to "think through" relevance decisions, making it effective for nuanced topics.

For specific benchmark results and comparisons with other models, please refer to the paper and the official GitHub repository.

Installation

Please see the Github for detailed installation instructions.

MTEB Integration

rank1 is compatible with the MTEB benchmarking framework:

from mteb import MTEB
from rank1 import rank1  # From the official repo

# Initialize the model
model = rank1(
    model_name_or_path="jhu-clsp/rank1-14b-awq",
    num_gpus=1,
    device="cuda",
    quantized=True  # Indicate that you're using the quantized version
)

# Run evaluation on specific tasks
evaluation = MTEB(tasks=["NevIR"])
results = evaluation.run(model)

Citation

If you use rank1 in your research, please cite our work:

@misc{weller2025rank1testtimecomputereranking,
      title={Rank1: Test-Time Compute for Reranking in Information Retrieval}, 
      author={Orion Weller and Kathryn Ricci and Eugene Yang and Andrew Yates and Dawn Lawrie and Benjamin Van Durme},
      year={2025},
      eprint={2502.18418},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2502.18418}, 
}

License

MIT License