Edit model card
YAML Metadata Warning: empty or missing yaml metadata in repo card (https://huggingface.co/docs/hub/model-cards#model-card-metadata)

PRMs are trained to predict the correctness of each step on the positions of "\n\n" and "<eos>".


import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "ScalableMath/llemma-7b-prm-prm800k-level-1to3-hf"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto")

tokenizer = AutoTokenizer.from_pretrained("EleutherAI/llemma_7b")

qa_example = """# Question

Convert the point $(0,3)$ in rectangular coordinates to polar coordinates.  Enter your answer in the form $(r,\theta),$ where $r > 0$ and $0 \le \theta < 2 \pi.$

# Solution

To convert from rectangular to polar coordinates, I need to use the formulas $r = \sqrt{x^2 + y^2}$ and $\theta = \tan^{-1}(y/x).$

In this case, $x = 0$ and $y = 3,$ so I can plug them into the formulas.

For $r,$ I get $r = \sqrt{0^2 + 3^2} = \sqrt{9} = 3.$

For $\theta,$ I get $\theta = \tan^{-1}(3/0).$

This is undefined, since the tangent function is not defined at $0.$

However, I can use the fact that the point $(0,3)$ lies on the positive $y$-axis, which has an angle of $\pi/2$ radians or $90^\circ.$

Therefore, I can choose any angle in the range $(0,\pi/2)$ as the value of $\theta.$

I will choose $\theta = \pi/2,$ since it is the simplest and most natural choice.

Therefore, the polar coordinates of the point $(0,3)$ are $(3,\pi/2).$

# Answer


begin_solution_tokens = tokenizer.encode("\n\n# Solution", add_special_tokens=False)[1:]
scoring_tokens = tokenizer.encode("\n\n", add_special_tokens=False)[1:]
eos_token = tokenizer.eos_token_id

input_ids = tokenizer.encode(qa_example)

begin_solution_flag = False

candidate_positions = []

for start_idx in range(len(input_ids)):
    if tuple(input_ids[start_idx:start_idx+len(begin_solution_tokens)]) == tuple(begin_solution_tokens):
        begin_solution_flag = True

    if begin_solution_flag and tuple(input_ids[start_idx:start_idx+len(scoring_tokens)]) == tuple(scoring_tokens):

    if input_ids[start_idx] == eos_token:

# maybe delete the first and the second to last candidate_positions
# because they are "\n\n" after "# Solution" and after "# Answer"
del candidate_positions[0]
del candidate_positions[-2]

input_tensor = torch.tensor([input_ids])
candidate_positions = torch.tensor(candidate_positions)

with torch.no_grad():
    logits = model(input_tensor).logits
    scores =logits.mean(dim=-1)
    step_scores = scores[0][candidate_positions]
    step_probs = torch.sigmoid(step_scores)


# tensor([0.9632, 0.8615, 0.9689, 0.9787, 0.9627, 0.7813, 0.9241, 0.3504, 0.7241, 0.8634])
Downloads last month

Collection including ScalableMath/llemma-7b-prm-prm800k-level-1to3-hf