Trump-Forecaster

RL-Tuned gpt-oss-120b for Predicting Trump Administration Actions

We fine-tuned gpt-oss-120b with reinforcement learning to predict Trump administration actions. Trained on the WWTD-2025 dataset of 2,108 binary forecasting questions generated with the Lightning Rod SDK, Trump-Forecaster beats GPT-5 on held-out forecasting questions.

This repo contains a LoRA adapter (5.3 GB) for gpt-oss-120b. A standalone merge.py script is included to produce a full merged model if needed.

Dataset · Lightning Rod SDK · Future-as-Label paper · Outcome-based RL paper

---

Results

Evaluated on 682 held-out test questions under two conditions: with news context, and without context (question only). The no-context condition reveals whether the model knows what it doesn't know—untrained models project false confidence, while RL training fixes overconfidence.

Model	Brier (With Context)	BSS	Brier (No Context)	BSS	ECE (With Context)	ECE (No Context)
GPT-5	0.200	+0.14	0.258	-0.11	0.091	0.191
gpt-oss-120b (base)	0.213	+0.08	0.260	-0.12	0.111	0.190
Trump-Forecaster	0.194	+0.16	0.242	-0.04	0.079	0.164

Metrics

• Brier Score: Mean squared error between predicted probability and outcome (0 or 1). Lower is better. Brier Skill Score (BSS) expresses this as improvement over always predicting the base rate—positive means the model learned something useful beyond historical frequency.
• Expected Calibration Error (ECE): Measures whether predicted probabilities match actual frequencies. "70%" predictions should resolve "yes" 70% of the time. Lower is better.

---

Training

• Base model: openai/gpt-oss-120b (120B MoE, 5.1B active params, 128 experts Top-4)
• Method: GRPO with Brier score reward via Tinker
• LoRA rank: 32
• Learning rate: 4e-5
• Batch size: 32, group size 8
• Training steps: 50
• Max tokens: 16,384

---

Usage

This repo contains a LoRA adapter trained with Tinker. The adapter uses Tinker's module naming convention, so it requires a merge step before inference. A standalone merge.py script is included.

Merge into full model

pip install torch transformers safetensors tqdm huggingface-hub
python merge.py --output ./trump-forecaster-merged

This downloads the base model, dequantizes to bf16, applies the LoRA adapter, and saves the merged model.

Inference

import sglang as sgl

engine = sgl.Engine(
    model_path="./trump-forecaster-merged",
    tokenizer_path="openai/gpt-oss-120b",
    trust_remote_code=True,
    dtype="bfloat16",
    tp_size=2,
)

news_context = "... relevant news articles ..."

prompt = f"""You are a forecasting expert. Given the question and context below, predict the probability that the answer is "Yes".

Question: Will Trump impose 25% tariffs on all goods from Canada by February 1, 2025?

Context:
{news_context}

Respond with your reasoning, then give your final answer as a probability between 0 and 1 inside <answer></answer> tags."""

output = engine.generate(prompt, sampling_params={"max_new_tokens": 4096, "stop": ["</answer>"]})
print(output["text"])

---

Links

• Dataset: LightningRodLabs/WWTD-2025
• Training platform: Tinker
• Data generation: Lightning Rod SDK
• Future-as-Label paper: arxiv:2601.06336
• Outcome-based RL paper: arxiv:2505.17989