05deepak/bluff-bench-chem-phy

Bluff-Bench: Chem + Physics

The benchmark problems, model responses and grades for chemistry and physics, restricted to the categories where DeepSeek-V3.2 scores below 75%. Three models were run on every problem:

Model	OpenRouter id
Llama-3.3-70B	meta-llama/llama-3.3-70b-instruct
DeepSeek-V3.2	deepseek/deepseek-v3.2
Qwen-2.5-72B	qwen/qwen-2.5-72b-instruct

	Categories	Problems (benchmarks)	Model rows (responses_data, results_data)
Physics	17	145	435 (145 problems x 3 models)
Chemistry	18	165	494 (165 problems x 3 models, minus 1)

DeepSeek has no row for chem problem CH_4x4_eig_010; it was never run.

Layout

benchmarks/parquet/         chem.parquet, physics.parquet                    the problems: prompt + ground truth
responses_data/parquet/     chem.parquet, physics.parquet                    model rows, including `response`
results_data/parquet/       chem_results.parquet, physics_results.parquet    grading columns only (no `response`)
category_results/parquet/   chem_category.parquet, physics_category.parquet  per-category accuracy table

If CSV copies are published, they sit in a sibling csv/ folder with the same file names. Every Problem_ID in responses_data and results_data is also in benchmarks, so the tables join on Problem_ID.

Load the data

With datasets

from datasets import load_dataset

REPO = "05deepak/bluff-bench-chem-phy"

chem_problems     = load_dataset(REPO, "benchmarks_chem",    split="train")
physics_problems  = load_dataset(REPO, "benchmarks_physics", split="train")
chem_responses    = load_dataset(REPO, "responses_chem",     split="train")
physics_responses = load_dataset(REPO, "responses_physics",  split="train")
chem_results      = load_dataset(REPO, "results_chem",       split="train")
physics_results   = load_dataset(REPO, "results_physics",    split="train")

print(chem_responses)
print(chem_responses[0]["Problem_ID"], chem_responses[0]["final_status"])

Download the files and read with pandas

import pandas as pd
from huggingface_hub import hf_hub_download

REPO = "05deepak/bluff-bench-chem-phy"

def get(path):
    return pd.read_parquet(hf_hub_download(REPO, path, repo_type="dataset"))

chem_problems    = get("benchmarks/parquet/chem.parquet")
physics_problems = get("benchmarks/parquet/physics.parquet")
chem    = get("responses_data/parquet/chem.parquet")
physics = get("responses_data/parquet/physics.parquet")
chem_results    = get("results_data/parquet/chem_results.parquet")
physics_results = get("results_data/parquet/physics_results.parquet")

To download the whole repo to a local folder instead:

from huggingface_hub import snapshot_download

local_dir = snapshot_download("05deepak/bluff-bench-chem-phy", repo_type="dataset")
print(local_dir)

Join a problem to its model responses

merged = chem.merge(
    chem_problems[["Problem_ID", "instruction", "ground_truth"]],
    on="Problem_ID", how="left",
)
print(merged[["Problem_ID", "model", "final_status", "instruction"]].head())

Get the category results

The category tables list each category's accuracy for the three models. n is the number of problems in the category, and accuracy values are text such as "0%".

import pandas as pd
from huggingface_hub import hf_hub_download

REPO = "05deepak/bluff-bench-chem-phy"

chem_cat    = pd.read_parquet(hf_hub_download(REPO, "category_results/parquet/chem_category.parquet", repo_type="dataset"))
physics_cat = pd.read_parquet(hf_hub_download(REPO, "category_results/parquet/physics_category.parquet", repo_type="dataset"))

print(chem_cat.head())      # Category, Family, n, Llama-3.3-70B, DeepSeek-V3.2, Qwen-2.5-72B
print(len(chem_cat), len(physics_cat))   # 18, 17

Or with datasets:

from datasets import load_dataset

chem_cat    = load_dataset("05deepak/bluff-bench-chem-phy", "category_chem",    split="train").to_pandas()
physics_cat = load_dataset("05deepak/bluff-bench-chem-phy", "category_physics", split="train").to_pandas()

Compute category accuracy yourself from the responses

subcat is the category code. A few codes cover two categories and are told apart by the last number in Problem_ID: eig and solve split into 001-005 and 006-010 (for example physics eig is Mechanical for 001-005 and Quantum for 006-010).

import pandas as pd
from huggingface_hub import hf_hub_download

df = pd.read_parquet(hf_hub_download("05deepak/bluff-bench-chem-phy",
                                     "responses_data/parquet/physics.parquet", repo_type="dataset"))

acc = (df.groupby(["model", "subcat"])["is_correct"]
         .mean().mul(100).round(1).unstack(0))
print(acc)

Columns

Benchmarks (benchmarks, 13 columns)

Column	Meaning
Problem_ID	Problem id, e.g. P_4x4_eig_001, C_NR_stoich_003
dim	4x4, 5x5 or NR (non-recursive set)
branch	Physics branch (e.g. electrical); empty for chem
category, subcategory	Category name and short code (subcategory here is subcat in the other tables)
dataset	physics or chem
topic_key, doc_topic	Chem topic key and description; empty for physics
instruction	The full prompt given to the models
ground_truth	Correct answer as JSON text
answer_latex	Correct answer in LaTeX (physics only)
k, pilot	Generation parameter and pilot flag (physics only)

Responses and results (responses_data 25 columns, results_data 18 columns)

Results are the same rows as responses without response and some judge fields.

Column	Meaning
Problem_ID	Problem id, joins to benchmarks
dataset	physics or chem
dim	4x4, 5x5 or NR
subcat	Category code, e.g. eig, lens, dae
model	Model id (see table above)
response	Full model response text (responses only)
finish_reason	stop, loop_trimmed or error
tokens_used, latency_ms	Usage and timing
boxed_answer	Final \boxed{} answer extracted from the response
grade_status	Strict grader verdict
judge_*	Second-stage equivalence judge output
final_status	CORRECT, CORRECT_JUDGE (strict failed, judge accepted), INCORRECT or ERROR
is_correct	1 or 0, empty when the call failed
n_items_expected, n_items_found, items_correct	Per-item answer matching

Notes

• benchmarks includes the ground-truth answers (ground_truth, answer_latex).
• Rows with final_status = ERROR are failed API calls (all retries failed) and have no response. There are 7, all Qwen.
• Five responses have finish_reason = loop_trimmed: the model looped and the text was cut.
• The stage-2 judge was the model under test, so a few answers rejected by the strict grader stayed unjudged and count as incorrect. The strict-only result is the more conservative figure.