Merge pull request #2 from PrimeIntellect-ai/feat/tmax-tasks

Add tmax task corpus

Author: rasdani

datasets/tmax/README.md

@@ -0,0 +1,5 @@
+# tmax
+
+TMax Harbor task corpus (`tmax@2026-07-01` in the research-environments registry), 14,600 tasks with prebuilt Prime container images (`prime/primeintellect/tmax:<task-dir>`).
+
+Moved from `PrimeIntellect-ai/research-environments` `environments/tmax_v1/tasks` at commit `e2064ba32` (PR #609 branch, includes the `task_004697_bd627cb1` numpy<2 build fix).

datasets/tmax/task_000000_c19dda5b/environment/Dockerfile

@@ -0,0 +1,15 @@
+FROM ubuntu:22.04
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV LC_ALL=C.UTF-8
+ENV LANG=C.UTF-8
+ENV GOPATH=/home/user/go
+ENV PATH=/home/user/go/bin:/home/user/.cargo/bin:$PATH
+
+COPY base_install.sh /tmp/base_install.sh
+RUN bash /tmp/base_install.sh && rm /tmp/base_install.sh
+
+COPY _fixtures/app/fixtures/image.png /app/fixtures/image.png
+
+COPY post_install.sh /tmp/post_install.sh
+RUN bash /tmp/post_install.sh && rm /tmp/post_install.sh

datasets/tmax/task_000000_c19dda5b/environment/_fixtures/app/fixtures/image.png

@@ -0,0 +1,69 @@
+apt-get update && apt-get install -y --no-install-recommends \
+        python3 python3-pip python3-venv \
+        coreutils findutils gawk sed grep \
+        curl wget git \
+        build-essential gcc g++ make cmake \
+        pkg-config autoconf automake libtool \
+        nodejs npm \
+        sqlite3 libsqlite3-dev \
+        valgrind gdb \
+        jq \
+        sudo \
+        golang-go \
+        rustc cargo \
+        nasm \
+        ffmpeg \
+        tesseract-ocr tesseract-ocr-eng \
+        upx-ucl binutils file xxd \
+        libjpeg-dev libpng-dev libtiff-dev libwebp-dev \
+        libfreetype6-dev liblapack-dev libopenblas-dev \
+        imagemagick \
+        espeak espeak-ng libespeak-ng-dev \
+        redis-server redis-tools \
+        nginx \
+        netcat-openbsd socat \
+        openmpi-bin libopenmpi-dev \
+        postgresql-client mariadb-client \
+        postfix mailutils \
+        expect inotify-tools \
+        unzip zip xz-utils \
+    && rm -rf /var/lib/apt/lists/*
+
+    # Core scientific Python stack — picks CPU-only torch via the dedicated
+    # PyPI index URL to avoid pulling the ~2 GB CUDA wheels.
+    #
+    # Includes a small set of common build / packaging / domain-specific
+    # helpers that v2 task deltas frequently invoke:
+    #   * pyinstaller         — compile Python → standalone binary
+    #     (caught by smoke: tasks asking the agent to package a CLI tool)
+    #   * cython              — C extension building from Python sources
+    #   * pybind11 / cffi     — C/C++ FFI for fixture-shipped binaries
+    #   * sympy               — symbolic math for metric_threshold verifiers
+    #   * networkx            — graph algorithms (debugging / scientific tasks)
+    #   * cryptography        — crypto / security tasks beyond pure stdlib
+    pip3 install --no-cache-dir \
+        pytest \
+        flask requests pyyaml toml \
+        numpy scipy scikit-learn pandas \
+        Pillow imageio matplotlib \
+        biopython \
+        beautifulsoup4 lxml \
+        pyinstaller cython pybind11 cffi \
+        sympy networkx cryptography \
+        redis psycopg2-binary pyserial paramiko
+
+    pip3 install --no-cache-dir \
+        --index-url https://download.pytorch.org/whl/cpu \
+        torch torchvision
+
+    # NOTE on the agent loop: rl_data.generator.vanillux_solver runs the
+    # mini-swe-agent-style harness ENTIRELY ON THE HOST (in our solver
+    # process). Only bash actions cross into this container via
+    # InteractiveContainerEnvironment.exec(). That means this SIF only
+    # needs bash + the build/ML/RE deps installed above — no swe-agent
+    # CLI, no Python 3.11 venv, no extra ~200 MB layer. As a bonus, the
+    # vanillux harness now also works on every legacy per-domain base
+    # SIF without modification.
+
+    useradd -m -s /bin/bash user || true
+    chmod 755 /home/user

datasets/tmax/task_000000_c19dda5b/environment/base_install.sh

@@ -0,0 +1,100 @@
+apt-get update && apt-get install -y python3 python3-pip git gcc make imagemagick tesseract-ocr fonts-dejavu-core
+pip3 install pytest
+
+mkdir -p /app
+convert -size 600x150 xc:white -font DejaVu-Sans -pointsize 24 -fill black -draw "text 10,40 'FATAL_OFFSET:0x1A4F\nTRACE: thread_worker -> extract_dump'" /app/bug_report.png
+
+useradd -m -s /bin/bash user || true
+mkdir -p /home/user/telemetry_repo
+cd /home/user/telemetry_repo
+git init
+git config user.email "dev@local"
+git config user.name "Dev"
+
+cat << 'EOF' > telemetry.c
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <pthread.h>
+
+char buffer[1024];
+pthread_mutex_t buffer_lock = PTHREAD_MUTEX_INITIALIZER;
+
+void* thread_worker(void* arg) {
+    char* input = (char*)arg;
+    pthread_mutex_lock(&buffer_lock);
+    strncpy(buffer, input, sizeof(buffer)-1);
+    buffer[sizeof(buffer)-1] = '\0';
+    pthread_mutex_unlock(&buffer_lock);
+    return NULL;
+}
+
+int main(int argc, char** argv) {
+    if (argc < 2) return 1;
+    pthread_t t1;
+    pthread_create(&t1, NULL, thread_worker, argv[1]);
+    pthread_join(t1, NULL);
+    printf("%s\n", buffer);
+    return 0;
+}
+EOF
+
+cat << 'EOF' > Makefile
+all:
+	gcc -o telemetry telemetry.c -lpthread
+EOF
+
+git add telemetry.c Makefile
+git commit -m "Commit 1"
+
+for i in $(seq 2 149); do
+    echo "// commit $i" >> telemetry.c
+    git commit -am "Commit $i"
+done
+
+make
+cp telemetry /app/oracle_parser
+strip /app/oracle_parser
+
+cat << 'EOF' > telemetry.c
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <pthread.h>
+
+char buffer[1024];
+pthread_mutex_t buffer_lock = PTHREAD_MUTEX_INITIALIZER;
+
+void* thread_worker(void* arg) {
+    char* input = (char*)arg;
+    strncpy(buffer, input, sizeof(buffer)-1);
+    buffer[sizeof(buffer)-1] = '\0';
+    return NULL;
+}
+
+int main(int argc, char** argv) {
+    if (argc < 2) return 1;
+    pthread_t t1;
+    pthread_create(&t1, NULL, thread_worker, argv[1]);
+    pthread_join(t1, NULL);
+    printf("%s\n", buffer);
+    return 0;
+}
+EOF
+
+cat << 'EOF' > Makefile
+all:
+	gcc -o telemetry telemetry.c
+EOF
+
+echo "// commit 150" >> telemetry.c
+git commit -am "Commit 150"
+
+for i in $(seq 151 200); do
+    echo "// commit $i" >> telemetry.c
+    git commit -am "Commit $i"
+done
+
+chown -R user:user /home/user/telemetry_repo
+chmod -R 777 /home/user
+chmod -R 777 /app

datasets/tmax/task_000000_c19dda5b/environment/post_install.sh

@@ -0,0 +1,13 @@
+You are tasked with resolving a critical concurrency regression in our C-based multithreaded telemetry parser. 
+
+A bug was recently introduced in our repository, causing race conditions that lead to memory corruption and invalid string extraction from system memory dumps. A user submitted a screenshot of the kernel panic and stack trace, but they didn't specify which commit caused it. 
+
+Here is your multi-step mission:
+
+1. **Analyze the Bug Report**: An image artifact is located at `/app/bug_report.png`. Use OCR (e.g., `tesseract`) to read the image. It contains a hidden configuration string and a specific memory offset (formatted like `FATAL_OFFSET:0x...`) that you will need to properly configure the lock boundaries in the code.
+2. **Bisect the Regression**: Navigate to the git repository at `/home/user/telemetry_repo`. There are exactly 200 commits. The earlier commits work flawlessly, but a recent commit introduced a race condition in `telemetry.c` and a subtle build configuration failure in the `Makefile`. Use bisection to find the offending commit.
+3. **Diagnose and Fix**: Once you identify the bad commit, analyze the concurrency bug (a missing mutex lock around the string extraction buffer). Fix the race condition in `telemetry.c` and resolve the build failure.
+4. **Compile the Target**: Your final, fixed C program must be compiled and placed at exactly `/home/user/fixed_parser`. 
+
+**Verification:**
+Your compiled executable `/home/user/fixed_parser` will be subjected to rigorous fuzz testing. It must be bit-exact equivalent in its standard output to our reference oracle binary for thousands of randomly generated memory dump strings. The executable should accept a single command-line argument (the memory string) and print the safely extracted output.

datasets/tmax/task_000000_c19dda5b/instruction.md

@@ -0,0 +1,28 @@
+schema_version = "1.1"
+
+[task]
+name = "tmax/task_000000_c19dda5b"
+description = "developer bisecting a regression across 200 commits"
+
+[metadata]
+source = "v2"
+base_image = "intricate"
+domain = "debugging"
+skill_type = "Algorithmic"
+primitive_skills = ["Memory dump analysis and string extraction", "Race condition and concurrency debugging", "Build failure diagnosis", "Fuzz testing"]
+task_complexity = "intricate task (multi-stage workflow combining (a) environment setup or package configuration, (b) primary implementation across multiple files or languages, (c) iterative refinement against a quantitative or adversarial verifier, and (d) a final integration step. Expect 30-60 commands.)"
+command_complexity = "bash, code, and system services (shell commands, scripts, package installation, service configuration, networking, and containers)"
+scenario = "developer bisecting a regression across 200 commits"
+language = "C"
+
+[agent]
+timeout_sec = 600.0
+
+[verifier]
+timeout_sec = 120.0
+
+[environment]
+docker_image = "prime/primeintellect/tmax:task_000000_c19dda5b"
+cpus = 1
+memory_mb = 2048
+allow_internet = true

datasets/tmax/task_000000_c19dda5b/task.toml

@@ -0,0 +1,16 @@
+#!/bin/bash
+set -e
+
+mkdir -p /logs/verifier
+
+cd /tests
+python3 -m pytest test_final_state.py -v 2>&1 | tee /logs/verifier/test-stdout.txt
+TEST_EXIT=${PIPESTATUS[0]}
+
+if [ $TEST_EXIT -eq 0 ]; then
+    echo 1 > /logs/verifier/reward.txt
+else
+    echo 0 > /logs/verifier/reward.txt
+fi
+
+exit 0

datasets/tmax/task_000000_c19dda5b/tests/test.sh

@@ -0,0 +1,69 @@
+# test_final_state.py
+
+import os
+import random
+import string
+import subprocess
+import pytest
+
+def test_fixed_parser_exists_and_executable():
+    path = "/home/user/fixed_parser"
+    assert os.path.exists(path), f"Expected file {path} does not exist. Did you compile your fixed C program?"
+    assert os.path.isfile(path), f"Expected {path} to be a file."
+    assert os.access(path, os.X_OK), f"Expected {path} to be executable."
+
+def test_oracle_parser_exists():
+    path = "/app/oracle_parser"
+    assert os.path.exists(path), f"Reference oracle {path} is missing."
+    assert os.path.isfile(path), f"Reference oracle {path} is not a file."
+    assert os.access(path, os.X_OK), f"Reference oracle {path} is not executable."
+
+def test_fuzz_equivalence():
+    agent_bin = "/home/user/fixed_parser"
+    oracle_bin = "/app/oracle_parser"
+
+    if not os.path.exists(agent_bin) or not os.access(agent_bin, os.X_OK):
+        pytest.fail(f"Cannot run fuzzing: {agent_bin} is missing or not executable.")
+    if not os.path.exists(oracle_bin) or not os.access(oracle_bin, os.X_OK):
+        pytest.fail(f"Cannot run fuzzing: {oracle_bin} is missing or not executable.")
+
+    random.seed(42)
+    charset = string.ascii_letters + string.digits + string.punctuation + " "
+
+    N = 5000
+
+    for i in range(N):
+        length = random.randint(10, 256)
+        test_input = "".join(random.choices(charset, k=length))
+
+        try:
+            oracle_proc = subprocess.run(
+                [oracle_bin, test_input],
+                capture_output=True,
+                text=True,
+                timeout=2
+            )
+            oracle_out = oracle_proc.stdout
+        except subprocess.TimeoutExpired:
+            pytest.fail(f"Oracle timed out on input: {test_input!r}")
+        except Exception as e:
+            pytest.fail(f"Oracle failed to run on input: {test_input!r}. Error: {e}")
+
+        try:
+            agent_proc = subprocess.run(
+                [agent_bin, test_input],
+                capture_output=True,
+                text=True,
+                timeout=2
+            )
+            agent_out = agent_proc.stdout
+        except subprocess.TimeoutExpired:
+            pytest.fail(f"Your parser timed out on input: {test_input!r}")
+        except Exception as e:
+            pytest.fail(f"Your parser failed to run on input: {test_input!r}. Error: {e}")
+
+        assert agent_out == oracle_out, (
+            f"Output mismatch on input: {test_input!r}\n"
+            f"Expected (Oracle): {oracle_out!r}\n"
+            f"Got (Agent): {agent_out!r}"
+        )

datasets/tmax/task_000000_c19dda5b/tests/test_final_state.py

@@ -0,0 +1,6 @@
+FROM ubuntu:22.04
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+COPY post_install.sh /tmp/post_install.sh
+RUN bash /tmp/post_install.sh && rm /tmp/post_install.sh