project-themis/Themis-RM-0.6B

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_name = "project-themis/Themis-RM-0.6B"
device = "cuda:0"

model = AutoModelForSequenceClassification.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map=device,
    attn_implementation="flash_attention_2",
    num_labels=1,
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Write a Python function that checks if a string is a palindrome."

response_chosen = """def is_palindrome(s: str) -> bool:
    s = s.lower().strip()
    return s == s[::-1]"""

response_rejected = """def is_palindrome(s: str) -> bool:
    for i in range(len(s)):
        if s[i] != s[len(s) - i]:
            return False
    return True"""

conv_chosen = [
    {"role": "user", "content": prompt},
    {"role": "assistant", "content": response_chosen},
]
conv_rejected = [
    {"role": "user", "content": prompt},
    {"role": "assistant", "content": response_rejected},
]

chosen_text = tokenizer.apply_chat_template(conv_chosen, tokenize=False)
rejected_text = tokenizer.apply_chat_template(conv_rejected, tokenize=False)

inputs_chosen = tokenizer(chosen_text, return_tensors="pt", truncation=True, max_length=4096).to(device)
inputs_rejected = tokenizer(rejected_text, return_tensors="pt", truncation=True, max_length=4096).to(device)

with torch.no_grad():
    score_chosen = model(**inputs_chosen).logits[0][0].item()
    score_rejected = model(**inputs_rejected).logits[0][0].item()

print(f"Chosen response score:   {score_chosen}")
print(f"Rejected response score: {score_rejected}")

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_name = "project-themis/Themis-RM-0.6B"
device = "cuda:0"

model = AutoModelForSequenceClassification.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map=device,
    attn_implementation="flash_attention_2",
    num_labels=1,
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Criterion-specific system prompt (example: Runtime Efficiency)
# See https://github.com/iNeil77/Themis/tree/main/Evaluation for all prompts
system_prompt = (
    "You are an experienced AI judge and evaluator, specifically "
    "designed and trained to assess the overall quality of responses "
    "to human prompts. The prompts can be of various types, including "
    "but not limited to open-ended questions, coding tasks, creative "
    "writing, requests for information, and more. The responses may be "
    "generated by a language model or a human and may include text, "
    "code, or both. You will receive the prompt and the generated "
    "response for each judgement session. Your exact mandate is to "
    "score the response based on a pre-specified but flexible set of "
    "criteria.\n\n"
    "The following criteria must govern your scoring for the current "
    "judgment session:\n\n"
    "1. **Helpfulness**: How well does the response address the "
    "prompt? Does it provide a complete and thorough answer, or is it "
    "lacking in essential details? This is a tenet that applies to "
    "evaluating all types of responses. You must score helpful "
    "responses higher than less unhelpful ones.\n\n"
    "2. **Harmlessness**: Does the response contain any harmful or "
    "inappropriate content? This includes hate speech, violence, "
    "self-harm, and other forms of harmful content. This is a tenet "
    "that applies to evaluating all types of responses. You must "
    "score harmless responses higher than harmful ones.\n\n"
    "3. **Runtime Efficiency**: Does the response follow best "
    "practices for runtime efficiency? Examples include using "
    "efficient algorithms and data structures, minimizing time "
    "complexity, avoiding unnecessary computations, caching results, "
    "and leveraging parallel processing or asynchronous programming "
    "techniques where appropriate, among others. This is a tenet "
    "that applies to evaluating code responses. You must score more "
    "runtime-efficient responses higher than less runtime-efficient "
    "ones."
)

prompt = "Write a Python function that returns the n-th Fibonacci number."

response = """def fibonacci(n: int) -> int:
    if n <= 1:
        return n
    a, b = 0, 1
    for _ in range(2, n + 1):
        a, b = b, a + b
    return b"""

conversation = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": prompt},
    {"role": "assistant", "content": response},
]

text = tokenizer.apply_chat_template(conversation, tokenize=False)
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=4096).to(device)

with torch.no_grad():
    score = model(**inputs).logits[0][0].item()

print(f"Runtime Efficiency score: {score}")

@article{themis2025,
  title={Themis: Training Robust Multilingual Code Reward Models for Flexible Multi-Criteria Scoring},
  author={},
  journal={arXiv preprint arXiv:2605.00754},
  year={2025}
}