forward-looking-examples

Forward-Looking Training Examples (GRPO)

---

Common Structure

Seeds → ForwardLookingQuestionGenerator → Labels → Context (optional) → Lint → Split → Train (GRPO)

from datetime import datetime
from lightningrod import (
    LightningRod, GRPOTrainingConfig, BinaryAnswerType,
    NewsSeedGenerator, ForwardLookingQuestionGenerator,
    NewsContextGenerator, WebSearchLabeler, QuestionPipeline,
    filter_and_split, display_lint_overview, get_lint_affected_sample_ids,
)
lr = LightningRod(api_key=api_key)

Default training config (Foresight uses num_rollouts=4 — large general datasets need fewer):

batch_size = 32
config = GRPOTrainingConfig(
    base_model_id="openai/gpt-oss-120b",
    training_steps=len(train_data) // batch_size,
    lora_rank=32,
    batch_size=batch_size,
    num_rollouts=8,
    max_response_length=16384,
    learning_rate=4e-5,
)

Lint the dataset — run on the full generated dataset before splitting. Linting runs server-side on the whole dataset, not on local subsets:

lint_result = lr.datasets.linter.run(dataset.id)
display_lint_overview(lint_result)

bad_ids = get_lint_affected_sample_ids(lint_result)
if bad_ids:
    clean_ids = [s.id for s in dataset.samples() if s.id not in set(bad_ids)]
    dataset = dataset.subset(clean_ids)

---

Example 1: Golf Forecasting

Goal: Predict professional golf outcomes — tournament winners, cuts, matchups, milestones.

Binary questions, news seeds, 14-day interval matching tournament cadence.

Source: lightningrod-python-sdk/notebooks/fine_tuning/01_golf_forecasting.ipynb (demo, max_seeds=20), llm_forecasting/notebooks/golf/golf.ipynb (production, 2000)

instructions = """
Generate binary forecasting questions about professional golf across all major tours and events.
Cover tournament outcomes, cuts, matchups, majors, team events, season races, world rankings,
and player milestones. Specific, verifiable, spanning the full probability spectrum.
"""

good_examples = [
    "Will Scottie Scheffler win the 2025 Masters?",
    "Will Tiger Woods make the cut at the 2025 Masters?",
    "Will any LIV player win a major championship in 2025?",
    "Will Europe win the 2025 Ryder Cup?",
    "Will Nelly Korda win the 2025 US Women's Open?",
]

bad_examples = [
    "Will someone win the tournament? (obvious)",
    "Will golf be exciting? (subjective)",
    "Will there be birdies? (trivial)",
]

pipeline = QuestionPipeline(
    seed_generator=NewsSeedGenerator(
        start_date=datetime(2024, 6, 1),       # base model knowledge cutoff
        end_date=datetime(2026, 1, 1),
        interval_duration_days=14,              # biweekly — tournament cadence
        search_query=[
            "PGA Tour", "LIV Golf", "LPGA", "golf major championship",
            "Ryder Cup Presidents Cup", "golf world rankings",
            "professional golf", "women's golf", "European Tour golf",
        ],
        articles_per_search=10,
    ),
    question_generator=ForwardLookingQuestionGenerator(
        instructions=instructions,
        examples=good_examples,
        bad_examples=bad_examples,
        answer_type=BinaryAnswerType(),
        questions_per_seed=5,
    ),
    context_generators=[
        NewsContextGenerator(articles_per_query=3, num_search_queries=3, num_articles=5)
    ],
    labeler=WebSearchLabeler(answer_type=BinaryAnswerType()),
)

# 2,000 is a good default — at questions_per_seed=5 that yields ~10,000 questions before filtering.
# Use max_seeds=20 to test the pipeline first.
dataset = lr.transforms.run(pipeline, max_seeds=2000, name="Golf forecasting")

train_dataset, test_dataset = filter_and_split(
    dataset, test_size=0.2, split_strategy="temporal",
    days_to_resolution_range=(1, None),
)

Train	Model	Brier	vs gpt-5
3,178	gpt-oss-120b	0.2074	17% better (BSS)

---

Example 2: Trump Policy Forecasting

Goal: Predict Trump administration policy actions, decisions, and political outcomes within a 2-month horizon.

Weekly intervals, high questions_per_seed (20) for a narrow domain.

Source: lightningrod-python-sdk/notebooks/fine_tuning/02_trump_forecasting.ipynb (demo, max_seeds=25), llm_forecasting/notebooks/e2e/wwtd_2025.ipynb (production)

instructions = """
Generate binary forecasting questions about Trump's actions, decisions, positions, and statements.
Diverse, covering the full range from very likely to very unlikely.
Horizon: outcomes known within 2 months. Binary, exact dates, self-contained, verifiable, newsworthy.
"""

good_examples = [
    "Will Trump impose 25% tariffs on all goods from Canada by February 1, 2025?",
    "Will Trump issue pardons to January 6 defendants within his first week in office?",
    "Will Pete Hegseth be confirmed as Secretary of Defense by February 15, 2025?",
    "Will Trump sign an executive order to keep TikTok operational in the US by January 31, 2025?",
    "Will Kash Patel be confirmed as FBI Director by March 1, 2025?",
]

bad_examples = [
    "Will Trump do something controversial? (too vague)",
    "Will Trump be in the news? (obvious)",
    "Will tariffs be imposed? (needs specifics)",
]

pipeline = QuestionPipeline(
    seed_generator=NewsSeedGenerator(
        start_date=datetime(2025, 1, 1),
        end_date=datetime(2026, 1, 1),
        interval_duration_days=7,               # weekly — fast-moving domain
        search_query=[
            "Donald Trump domestic policy agenda",
            "Donald Trump trade and tariff actions",
            "Donald Trump foreign policy decisions",
            "Donald Trump interviews and press appearances",
            "Donald Trump lawsuits and court rulings",
        ],
        articles_per_search=10,
    ),
    question_generator=ForwardLookingQuestionGenerator(
        instructions=instructions,
        examples=good_examples,
        bad_examples=bad_examples,
        answer_type=BinaryAnswerType(),
        questions_per_seed=20,                  # narrow domain, many angles per article
    ),
    context_generators=[
        NewsContextGenerator(articles_per_query=3, num_search_queries=1, num_articles=5)
    ],
    labeler=WebSearchLabeler(answer_type=BinaryAnswerType()),
)

train_dataset, test_dataset = filter_and_split(
    dataset, test_size=0.2, split_strategy="temporal",
    days_to_resolution_range=(1, 60),           # 2-month horizon
)

Train	Model	Brier	vs gpt-5
2,108	gpt-oss-120b	0.1939	Beat (0.2003)

---

Example 3: Military Strikes (Large Scale)

Goal: Predict military strike events — airstrikes, missile strikes, drone strikes, naval operations — across global conflicts.

Large-scale generation (10k questions). Detailed instructions requiring specific actor/target naming — vague military questions are useless.

Source: lightningrod-python-sdk/notebooks/fine_tuning/04_military_strikes.ipynb

instructions = """
Generate binary forecasting questions about military strikes and attack operations.

Cover all strike types: airstrikes, missile strikes, drone strikes, naval strikes.
Cover state and non-state actors using natural news language.

Questions must name the specific actor, specific target, have a specific date/milestone,
be verifiable from open-source news, and span the full probability spectrum.
"""

good_examples = [
    "Will the IDF conduct airstrikes on Hezbollah weapons depots in the Bekaa Valley before November 2024?",
    "Will Iran launch a direct ballistic missile strike on Israeli territory before May 2024?",
    "Will Ukraine conduct drone strikes on Russian oil refineries before April 2024?",
    "Will the US Navy conduct Tomahawk strikes on Houthi radar sites before February 2024?",
    "Will NATO aircraft conduct strikes inside Russian territory before January 2025?",
]

bad_examples = [
    "Will there be an attack somewhere? (no actor, target, or location)",
    "Will violence increase in the Middle East? (vague)",
    "Will conflict continue in Ukraine? (trivially obvious)",
    "Will missiles be fired? (no actor, no target, no date)",
]

pipeline = QuestionPipeline(
    seed_generator=NewsSeedGenerator(
        start_date=datetime(2024, 6, 1),
        end_date=datetime(2026, 3, 1),
        interval_duration_days=7,
        search_query=[
            "military airstrike", "military strike",
            "missile strike", "drone strike", "naval strike",
        ],
        articles_per_search=10,
    ),
    question_generator=ForwardLookingQuestionGenerator(
        instructions=instructions,
        examples=good_examples,
        bad_examples=bad_examples,
        answer_type=BinaryAnswerType(),
        questions_per_seed=5,
    ),
    context_generators=[
        NewsContextGenerator(articles_per_query=3, num_search_queries=3, num_articles=5)
    ],
    labeler=WebSearchLabeler(answer_type=BinaryAnswerType()),
)

dataset = lr.transforms.run(pipeline, max_seeds=2000, name="Military strikes forecasting")

train_dataset, test_dataset = filter_and_split(
    dataset, test_size=0.2, split_strategy="temporal",
    days_to_resolution_range=(1, 90),
)

batch_size = 32
config = GRPOTrainingConfig(
    base_model_id="openai/gpt-oss-120b",
    training_steps=len(train_dataset.flattened()) // batch_size,
    lora_rank=32, batch_size=batch_size, num_rollouts=8,
    max_response_length=16384, learning_rate=4e-5,
)
job = lr.training.run(config, dataset=train_dataset, name="military-strikes-v1")

from lightningrod import EvalModel

eval_job = lr.evals.run_from_training_job(
    config,
    job,
    test_dataset,
    # Frontier benchmark — always included by default (see lightningrod-assistant).
    extra_models=[EvalModel(model_id="openai/gpt-5.5", label="GPT-5.5")],
)

# Optional: if eval scores are disappointing, re-run with reasoning comparison
# to see HOW the base and fine-tuned models reason differently:
# eval_job = lr.evals.run_from_training_job(
#     config, job, test_dataset,
#     reasoning_comparison_sample_size=20,
# )
# eval_job.reasoning_comparison_report

---

Example 4: Foresight v3 — General Forecasting (GDELT)

Goal: Build a general-purpose forecasting model across all domains — politics, economics, science, geopolitics.

Seeds from GDELT (broad global news). High confidence threshold (0.9) for production quality. Custom answer format.

Source: llm_forecasting/notebooks/foresight-gpt-oss/data-generation-binary.ipynb, tinker-train-with-template.ipynb

from lightningrod import GdeltSeedGenerator, QuestionRenderer

instructions = """
Generate forward-looking binary forecasting questions with exactly one yes/no answer.
Must be unresolved when asked, resolve within 3 months, materially important, fully self-contained,
and resolvable via public web search. Do NOT include numeric, multiple-choice, trivial, vague,
or > 3-month questions.
"""

good_examples = [
    "Will the U.S. Federal Reserve cut interest rates at its FOMC meeting ending May 6, 2026?",
    "Will the S&P 500 index close above 5,000 on April 30, 2026?",
    "Will the Liberal Party win a majority in the Canadian federal election on April 28, 2026?",
    "Will NASA's Artemis II crewed lunar flyby mission launch before June 30, 2026?",
    "Will Brent crude oil futures close above $80 per barrel on June 30, 2026?",
]

bad_examples = [
    "What will the inflation rate be in March 2026? # numeric, not binary",
    "Will inflation rise? # no time frame, no metric",
    "Which party will win the election? # multiple-choice",
    "Will the economy improve next year? # vague, exceeds 3 months",
]

answer_type = BinaryAnswerType(
    answer_format_instruction=(
        "Think carefully about your answer and output your final prediction "
        "(a float between 0.0 and 1.0) between <answer></answer> tags."
    ),
)

pipeline = QuestionPipeline(
    seed_generator=GdeltSeedGenerator(
        start_date=datetime(2024, 7, 3),
        end_date=datetime(2025, 11, 30),
        interval_duration_days=7,
        # Higher than NewsSeedGenerator — GDELT covers broad global topics,
        # so more articles per interval still yields diverse seeds.
        articles_per_interval=25,
    ),
    question_generator=ForwardLookingQuestionGenerator(
        instructions=instructions,
        examples=good_examples,
        bad_examples=bad_examples,
        questions_per_seed=5,
        answer_type=answer_type,
    ),
    context_generators=[NewsContextGenerator(num_articles=10)],
    labeler=WebSearchLabeler(answer_type=answer_type, confidence_threshold=0.9),
    renderer=QuestionRenderer(answer_type=answer_type),
)

dataset = lr.transforms.run(pipeline, max_seeds=1800)

Training: gpt-oss-120b, num_rollouts=4 (sufficient for large general dataset), training_steps=160.

---

Example 5: Timestamped Documents (FileSet RAG)

Goal: Train a domain expert from a corpus of timestamped documents, with no pre-existing labeled data.

General pattern for any set of documents that arrive over time — quarterly reports, research papers, regulatory filings, internal memos, etc. Generate questions from earlier documents, resolve them with later ones via FileSetRAGLabeler with TemporalConstraint.AFTER.

This example uses Fed Beige Book reports, but the pattern applies to any corpus where later documents can confirm or deny outcomes discussed in earlier ones.

Source: llm_forecasting/notebooks/filesets/beigebook/01_beige_book_pipeline.ipynb (branch: beige-book-e2e)

from lightningrod import FileSetSeedGenerator

# Upload documents to FileSet with date metadata.
# Each document needs a date so the labeler can enforce temporal ordering.
# (see source notebook for download + upload code)

pipeline = QuestionPipeline(
    seed_generator=FileSetSeedGenerator(
        file_set_id=fileset.id, chunk_size=4000, chunk_overlap=200,
    ),
    question_generator=ForwardLookingQuestionGenerator(
        questions_per_seed=5,
        answer_type=BinaryAnswerType(),
        instructions=(
            "Generate questions about specific economic outcomes (decrease, increase, slow, "
            "accelerate). Ask about the outcome directly — 'Will loan nonperformance in Dallas "
            "decrease?' — NOT 'Will the next Beige Book report that...'. "
            "No dates/months/years. Focus on district-specific topics. "
            "Flat/stable/unchanged = No. Not reported = Undetermined."
        ),
        examples=[
            "Will loan nonperformance in the Dallas district decrease?",
            "Will employment growth in the Philadelphia district slow?",
            "Will manufacturing activity in the St. Louis district improve?",
        ],
        bad_examples=[
            "Will the next Beige Book report that X? # report framing",
            "Will the October 2024 Beige Book report X? # specific dates",
            "Will economic activity in Cleveland change? # too vague",
        ],
    ),
)

dataset = lr.transforms.run(pipeline, max_seeds=600)

# Phase 2: Label via FileSetRAGLabeler with TemporalConstraint.AFTER
# This ensures only documents published AFTER the question's seed date are used for labeling,
# so the model can't "see" the answer at question time.
# confidence_threshold=0.7 (lower — answers from known documents)
# labeler_instruction emphasizes: flat/stable = No, not reported = Undetermined,
# never infer No from absence

Training: gpt-oss-120b, training_steps=66, num_rollouts=4.

---

Reference

filter_and_split()

Temporal splitting ensures no training sample has a close date past the start of the test set, preventing data leakage. days_to_resolution_range filters to resolved questions within the expected horizon.

train_dataset, test_dataset = filter_and_split(
    dataset, test_size=0.2, split_strategy="temporal",
    days_to_resolution_range=(1, None),  # resolved only
)

Domain	days_to_resolution_range
Politics	(1, 60)
Sports	(1, None)
Geopolitics	(1, 90)
General	(1, None)

Training Parameters

Parameter	Value	Notes
base_model_id	openai/gpt-oss-120b	Default
lora_rank	32 or 16	Standard
batch_size	32	Standard
num_rollouts	8	More = better signal
max_response_length	16384	Reasoning chains
learning_rate	4e-5	Standard
training_steps	len(train_data) // batch_size	One pass

Cost Estimation

cost = lr.training.estimate_cost(config, dataset=train_dataset)