File size: 3,500 Bytes
4bc0c95 7041780 4bc0c95 7041780 4bc0c95 7041780 4bc0c95 7041780 4bc0c95 7041780 4bc0c95 7041780 4bc0c95 7041780 4bc0c95 7041780 4bc0c95 7041780 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 | ---
license: apache-2.0
task_categories:
- question-answering
- text-generation
tags:
- reasoning
- evaluation
- math
- benchmark-results
- model-outputs
size_categories:
- 100K<n<1M
---
# Forecast-Generalization: raw evaluation outputs across 38 reasoning models
Complete generation-level outputs, per-seed scores and analysis artifacts from a
study of how well benchmark performance **forecasts generalization** to held-out
reasoning tasks.
Most released evaluations report only aggregate accuracy. This release keeps the
**raw per-problem, per-seed generations**, so item-level analyses can be redone
without re-running any inference.
## What is here
**38 models** evaluated on **four competition-style reasoning benchmarks**, each
run with **five random seeds** (42, 123, 456, 789, 1024).
| benchmark | evaluation runs |
|---|---|
| HMMT | 36 |
| GPQA | 30 |
| AIME | 30 |
| IMO | 22 |
Model families covered include OLMo-3 (7B/32B, Think and Think-SFT variants),
DeepSeek-R1 distillations (Llama-8B/70B, Qwen-7B/14B/32B), Qwen3, Gemma-4,
Seed-OSS-36B, Kimi, MiniMax and others — spanning roughly 7B to 70B parameters
plus several API-served models.
## Layout
| path | contents |
|---|---|
| `outputs/` | one directory per (benchmark, model, run): per-seed `*_predictions_seed-N.tsv` and `*_logs_seed-N.txt` |
| `expert_outputs/` | reference/expert model generations, grouped by `benchmark_model` |
| `scores/` | per-model accuracy CSVs: `name, seed_42 … seed_1024, mean_accuracy, std_error, num_seeds` |
| `data/` | benchmark inputs |
| `proxy_results/*.tar.gz` | proxy-metric evaluation results (next-token-prediction losses, sweeps) |
| `plots.tar.gz` | generated figures |
| `MANIFEST.tsv` | per-archive file counts, byte sizes and sha256 |
## Why the tarballs
`proxy_results/` and `plots/` together hold about 129,000 small JSON and image
files. Uploaded raw they would exceed practical per-repository file counts, so
they are stored as gzipped tars split along their natural subset boundaries.
Every archive was verified with `tar -d` (a member-by-member diff against the
source filesystem) after creation, and `MANIFEST.tsv` records the source file
count, archive member count, size and sha256 for each. Extraction restores the
original tree exactly:
```bash
tar -xzf proxy_results/ntp_results.tar.gz
```
## Example: seed variance on AIME
```python
import pandas as pd
from huggingface_hub import hf_hub_download
p = hf_hub_download("dvader13/forecastgen-artifacts",
"scores/aime_scores_seeds.csv", repo_type="dataset")
df = pd.read_csv(p)
print(df.sort_values("mean_accuracy", ascending=False)
[["name", "mean_accuracy", "std_error"]].head())
```
Each row is one model; the five `seed_*` columns give per-seed accuracy, so
seed-level variance is directly measurable rather than inferred.
## Intended uses
- Item-level analysis (which problems a model solves, not just how many)
- Seed-variance and evaluation-reliability studies
- Correlating cheap proxy metrics against downstream benchmark accuracy
- Cross-model comparison on identical problems and identical seeds
## Notes
- Generations are raw model output, unfiltered, and include incorrect answers —
that is deliberate, since error analysis is a main intended use.
- Benchmark problem statements belong to their original sources and retain the
licenses of those benchmarks; the license above applies to the outputs and
analysis artifacts collected here.
|