File size: 9,027 Bytes
6d5d207 4f8a73c 6d5d207 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 |
---
license: mit
datasets:
- jhu-clsp/rank1-training-data
base_model:
- jhu-clsp/rank1-llama3-8b
pipeline_tag: text-generation
tags:
- reranker
- retrieval
- quantized
- awq
language:
- en
---
# rank1-llama3-8b-awq: Quantized Model for Test-Time Compute Reranking
๐ [Paper](https://arxiv.org/abs/2502.18418) | ๐ [GitHub Repository](https://github.com/orionw/rank1)
rank1-llama3-8b-awq is a quantized version of the rank1-llama3-8b model. This AWQ-quantized 8B parameter model maintains the reasoning capabilities of the original model while requiring less memory and providing faster inference. The model is trained from the Llama 3.1 8B 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](https://huggingface.co/jhu-clsp/rank1-7b) | Qwen2.5-7B | Qwen variant (7B parameters) |
| [rank1-14b](https://huggingface.co/jhu-clsp/rank1-14b) | Qwen2.5-14B | Qwen variant (14B parameters) |
| [rank1-32b](https://huggingface.co/jhu-clsp/rank1-32b) | Qwen2.5-32B | Qwen variant (32B parameters) |
| [rank1-mistral-2501-24b](https://huggingface.co/jhu-clsp/rank1-mistral-2501-24b) | Mistral-Small 2501 24B | Mistral variant (24B parameters) |
| [rank1-llama3-8b](https://huggingface.co/jhu-clsp/rank1-llama3-8b) | Llama 3.1 8B | Full-precision version (8B parameters) |
### Quantized Variants
| Model | Description |
|:------|:------------|
| [rank1-7b-awq](https://huggingface.co/jhu-clsp/rank1-7b-awq) | Quantized version of rank1-7b |
| [rank1-14b-awq](https://huggingface.co/jhu-clsp/rank1-14b-awq) | Quantized version of rank1-14b |
| [rank1-32b-awq](https://huggingface.co/jhu-clsp/rank1-32b-awq) | Quantized version of rank1-32b |
| [rank1-mistral-2501-24b-awq](https://huggingface.co/jhu-clsp/rank1-mistral-2501-24b-awq) | Quantized version of rank1-mistral-24b |
| [rank1-llama3-8b-awq](https://huggingface.co/jhu-clsp/rank1-llama3-8b-awq) | Current model - Quantized version of rank1-llama3-8b |
## Associated Data and Resources
| Resource | Description |
|:---------|:------------|
| [rank1-r1-msmarco](https://huggingface.co/datasets/jhu-clsp/rank1-r1-msmarco) | All R1 output examples from MS MARCO |
| [rank1-training-data](https://huggingface.co/datasets/jhu-clsp/rank1-training-data) | Training data used for rank1 models |
| [rank1-run-files](https://huggingface.co/datasets/jhu-clsp/rank1-run-files) | Pre-computed run files for use in top 100 doc reranking |
| [GitHub Repository](https://github.com/orionw/rank1) | 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.
<details>
<summary>Click to expand: Minimal example with vLLM</summary>
```python
from vllm import LLM, SamplingParams
import math
# Initialize the model with vLLM
model = LLM(
model="jhu-clsp/rank1-llama3-8b-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-llama3-8b-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}")
```
</details>
<details>
<summary>Click to expand: Usage with AutoGPTQ/AWQ</summary>
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
# Load the tokenizer and quantized model
tokenizer = AutoTokenizer.from_pretrained("jhu-clsp/rank1-llama3-8b-awq")
model = AutoModelForCausalLM.from_pretrained(
"jhu-clsp/rank1-llama3-8b-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}")
```
</details>
## Performance
rank1-llama3-8b-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](https://github.com/embeddings-benchmark/mteb):
```python
from mteb import MTEB
from rank1 import rank1 # From the official repo
# Initialize the model
model = rank1(
model_name_or_path="jhu-clsp/rank1-llama3-8b-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:
```bibtex
@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](https://github.com/orionw/rank1/blob/main/LICENSE) |