SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2310.08491 · ICLR · 2024

Prometheus: Inducing Fine-grained Evaluation Capability in Language Models

Jamin Shin, Naver Lab, Minjoon Seo, James Thorne, Yun Sangdoo, Joel Jang, Shayne Longpre, Sungdong Kim, Seungone Kim, Yejin Cho, Hwaran Lee, Seongjin Shin

arXiv · PDF · Open in the Atlas

Code that ran

4 of 9 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 4 failures of our instrument; 5 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2310.08491.json · badge: /api/ran/2310.08491.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
kaistAI/Prometheus canonical (classifier label)
classifier basis: census
4 ran · 5 no recorded run
FunctionStatusWhere it lives
encode_with_messages_format Ran
ran; our contract draft was wrong, not the code
kaistAI/Prometheus/train/utils.py
pointer only (licence: NONE) · get_code("1cb0eeb16cf04ee3")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, kaistAI/Prometheus when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_next_word_predictions Ran
ran; our contract draft was wrong, not the code
kaistAI/Prometheus/train/utils.py
code served (permissive licence) · get_code("033263e69f60d9fe")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Minimal model that satisfies the function's usage
class MinimalLanguageModel(nn.Module):
    def __init__(self, vocab_size=100, hidden_dim=4):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, hidden_dim)
        self.linear = nn.Linear(hidden_dim, vocab_size)
        self.device_type = 'cpu'
    
    @property
    def device(self):
        class Device:
            type = 'cpu'
        return Device()
    
    def forward(self, input_ids, attention_mask=None):
        x = self.embedding(input_ids)
        x = x.mean(dim=1)
        logits = self.linear(x)
        
        class Output:
            pass
        
        out = Output()
        out.logits = logits.unsqueeze(1)  # Shape: (batch_size, 1, vocab_size)
        return out

# Minimal tokenizer stand-in
class MinimalTokenizer:
    def __init__(self, vocab_size=100):
        self.vocab_size = vocab_size
        self.vocab = {f"token_{i}": i for i in range(vocab_size)}
        self.id_to_token = {i: f"token_{i}" for i in range(vocab_size)}
    
    def __call__(self, texts, padding=None, return_tensors=None, add_special_tokens=True):
        if isinstance(texts, str):
            texts = [texts]
        
        batch_size = len(texts)
        max_len = 5
        
        input_ids = torch.randint(0, self.vocab_size, (batch_size, max_len))
        attention_mask = torch.ones_like(input_ids)
        
        class TokenizerOutput:
            pass
        
        output = TokenizerOutput()
        output.input_ids = input_ids
        output.attention_mask = attention_mask
        return output
    
    def convert_ids_to_tokens(self, ids):
        if isinstance(ids, torch.Tensor):
            ids = ids.tolist()
        if isinstance(ids, list):
            return [self.id_to_token.get(int(i), f"token_{i}") for i in ids]
        else:
            return self.id_to_token.get(int(ids), f"token_{ids}")

model = MinimalLanguageModel(vocab_size=100, hidden_dim=4)
tokenizer = MinimalTokenizer(vocab_size=100)
prompts = ["Hello world", "How are you"]

ARGS = {
    'model': model,
    'tokenizer': tokenizer,
    'prompts': prompts,
    'candidate_token_ids': None,
    'batch_size': 1,
    'return_token_predictions': False,
    'add_special_tokens': True,
    'disable_tqdm': True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, tqdm — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_llama_from_config Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
kaistAI/Prometheus/train/utils/model_utils.py
pointer only (licence: NONE); the stored copy get_code serves, from zhu-minjun/safetylock, is pointer only (licence: NONE) · get_code("b4c052fb373aef07")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, kaistAI/Prometheus when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

load_peft_model Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
kaistAI/Prometheus/train/utils/model_utils.py
pointer only (licence: NONE); the stored copy get_code serves, from zhu-minjun/safetylock, is pointer only (licence: NONE) · get_code("6d5ba8546b6aad2c")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, kaistAI/Prometheus when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

create Not yet run
unverified (harvested, no recorded run)
kaistAI/Prometheus/evaluation/benchmark/run_absolute_scoring.py
code served (permissive licence) · get_code("5743b6ea08e5b7b3")
How it ran

no invocation recorded.

create Not yet run
unverified (harvested, no recorded run)
kaistAI/Prometheus/evaluation/benchmark/run_relative_scoring.py
code served (permissive licence) · get_code("b953d7f234bcb588")
How it ran

no invocation recorded.

generate_completions Not yet run
unverified (harvested, no recorded run)
kaistAI/Prometheus/train/utils.py
pointer only (licence: NONE) · get_code("352416cd121404a7")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, kaistAI/Prometheus when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

load_model Not yet run
unverified (harvested, no recorded run)
kaistAI/Prometheus/train/utils/model_utils.py
code served (permissive licence) · get_code("2752e49a6edeae4e")
How it ran

no invocation recorded.

load_sharded_model_single_gpu Not yet run
unverified (harvested, no recorded run)
kaistAI/Prometheus/train/model_checkpointing/checkpoint_handler.py
pointer only (licence: NONE) · get_code("abd14de7398ccffe")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, kaistAI/Prometheus when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

Also present in other papers' code

Of 9 samples lifted from this paper's linked repositories, 9 carry a function identity and 0 do not. Of the 9 distinct functions with an identity, 5 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 4 the same month, 5 later) and 4 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
encode_with_messages_format 8 other papers
1 the same month, 7 later
2310.11511 (2023-10)
Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection
2311.05657 (2023-11)
Agent Lumos: Unified and Modular Training for Open-Source Language Agents
2401.08565 (2024-01)
Tuning Language Models by Proxy
and 5 more not named here
load_peft_model 7 other papers
1 dated earlier, 2 the same month, 4 later
2308.10855 (2023-08)
LatEval: An Interactive LLMs Evaluation Benchmark with Incomplete Information from Lateral Thinking Puzzles
2310.02469 (2023-10)
PrivacyMind: Large Language Models Can Be Contextual Privacy Protection Learners
2310.06498 (2023-10)
A New Benchmark and Reverse Validation Method for Passage-level Hallucination Detection
and 4 more not named here
generate_completions 6 other papers
6 later
2401.10768 (2024-01)
Knowledge Verification to Nip Hallucination in the Bud
2401.15006 (2024-01)
Airavata: Introducing Hindi Instruction-tuned LLM
2401.16405 (2024-01)
Scaling Sparse Fine-Tuning to Large Language Models
and 3 more not named here
load_llama_from_config 6 other papers
2 the same month, 4 later
2310.02469 (2023-10)
PrivacyMind: Large Language Models Can Be Contextual Privacy Protection Learners
2310.06498 (2023-10)
A New Benchmark and Reverse Validation Method for Passage-level Hallucination Detection
2403.04746 (2024-03)
LLMs in the Imaginarium: Tool Learning through Simulated Trial and Error
and 3 more not named here
load_sharded_model_single_gpu 4 other papers
1 the same month, 3 later
2310.02469 (2023-10)
PrivacyMind: Large Language Models Can Be Contextual Privacy Protection Learners
2402.12030 (2024-02)
Towards Cross-Tokenizer Distillation: the Universal Logit Distillation Loss for LLMs
2402.18540 (2024-02)
Keeping LLMs Aligned After Fine-tuning: The Crucial Role of Prompt Templates
and 1 more not named here

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Recently, using a powerful proprietary Large Language Model (LLM) (e.g., as an evaluator for long-form responses has become the de facto standard. However, for practitioners with large-scale evaluation tasks and custom criteria in consideration (e.g., child-readability), using proprietary LLMs as an evaluator is unreliable due to the closed-source nature, uncontrolled versioning, and prohibitive costs. In this work, we propose PROMETHEUS, a fully open-source LLM that is on par with GPT-4's evaluation capabilities when the appropriate reference materials (reference answer, score rubric) are accompanied. We first construct the FEEDBACK COLLECTION, a new dataset that consists of 1K fine-grained score rubrics, 20K instructions, and 100K responses and language feedback generated by GPT-4. Using the FEEDBACK COLLECTION, we train PROMETHEUS, a 13B evaluator LLM that can assess any given long-form text based on customized score rubric provided by the user. Experimental results show that PROMETHEUS scores a Pearson correlation of 0.897 with human evaluators when evaluating with 45 customized score rubrics, which is on par with GPT-4 (0.882), and greatly outperforms ChatGPT (0.392). Furthermore, measuring correlation with GPT-4 with 1222 customized score rubrics across four benchmarks (MT Bench, Vicuna Bench, Feedback Bench, Flask Eval) shows similar trends, bolstering PROMETHEUS's capability as an evaluator LLM. Lastly, PROMETHEUS achieves the highest accuracy on two human preference benchmarks (HHH Alignment & MT Bench Human Judgment) compared to open-sourced reward models explicitly trained on human preference datasets, highlighting its potential as an universal reward model. We open-source our code, dataset, and model 1 . * denotes equal contribution. Work was done while Seungone was interning at NAVER AI Lab.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2310.08491")
get_code_for_paper("2310.08491")
have("2310.08491")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2310.08491.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2310.08491.svg)](https://syntology.ai/paper/2310.08491)

Connect an agent — have() is free.