Add dataset card
Browse files
README.md
ADDED
|
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- baseweight-benchmark
|
| 5 |
+
- benchmark
|
| 6 |
+
- predictions
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
# Baseweight Benchmark — Raw Predictions
|
| 10 |
+
|
| 11 |
+
Every model generation produced during evaluation for the
|
| 12 |
+
[Baseweight Benchmark](https://baseweight.co/benchmark), a reproducible
|
| 13 |
+
comparison of fine-tuned open models against frontier APIs on focused vertical
|
| 14 |
+
tasks.
|
| 15 |
+
|
| 16 |
+
## What's here
|
| 17 |
+
|
| 18 |
+
One JSONL file per model × task × condition × seed. Each line is one
|
| 19 |
+
prediction:
|
| 20 |
+
|
| 21 |
+
| Field | Description |
|
| 22 |
+
|---|---|
|
| 23 |
+
| `id` | Row identifier from the prepared test split |
|
| 24 |
+
| `model` | Model short name (`qwen3-8b`, `gpt-5.4-mini`) |
|
| 25 |
+
| `condition` | `lora`, `zero-shot`, or `5-shot` |
|
| 26 |
+
| `eval_seed` | Seed used for this eval run (0, 1, or 2) |
|
| 27 |
+
| `prompt_sha` | SHA-256 prefix of the prompt template used |
|
| 28 |
+
| `input` | The text input sent to the model |
|
| 29 |
+
| `output` | The raw model output |
|
| 30 |
+
| `ground_truth` | Gold label or answer |
|
| 31 |
+
| `input_tokens` | Prompt token count |
|
| 32 |
+
| `output_tokens` | Completion token count |
|
| 33 |
+
| `avg_logprob` | Mean log-probability of output tokens |
|
| 34 |
+
| `timestamp` | UTC time of the generation |
|
| 35 |
+
|
| 36 |
+
## File layout
|
| 37 |
+
|
| 38 |
+
```
|
| 39 |
+
predictions/
|
| 40 |
+
local/
|
| 41 |
+
qwen3-8b/
|
| 42 |
+
banking77/ lora.jsonl lora_seed1.jsonl lora_seed2.jsonl
|
| 43 |
+
zero-shot.jsonl zero-shot_seed1.jsonl zero-shot_seed2.jsonl
|
| 44 |
+
5-shot.jsonl 5-shot_seed1.jsonl 5-shot_seed2.jsonl
|
| 45 |
+
cuad/ lora.jsonl lora_seed1.jsonl lora_seed2.jsonl
|
| 46 |
+
zero-shot.jsonl zero-shot_seed1.jsonl zero-shot_seed2.jsonl
|
| 47 |
+
api/
|
| 48 |
+
gpt-5.4-mini/
|
| 49 |
+
banking77/ zero-shot.jsonl zero-shot_seed1.jsonl zero-shot_seed2.jsonl
|
| 50 |
+
5-shot.jsonl 5-shot_seed1.jsonl 5-shot_seed2.jsonl
|
| 51 |
+
cuad/ zero-shot.jsonl zero-shot_seed1.jsonl zero-shot_seed2.jsonl
|
| 52 |
+
```
|
| 53 |
+
|
| 54 |
+
Seeds 1 and 2 are independent re-evaluations with shuffled test order.
|
| 55 |
+
Seed 0 is the base run (file without seed suffix). Reported metrics are the
|
| 56 |
+
mean across all three seeds; 95% CIs are bootstrap intervals over the seeds.
|
| 57 |
+
|
| 58 |
+
## Verifying the numbers
|
| 59 |
+
|
| 60 |
+
Download any JSONL, recompute the metric against `ground_truth`, and compare
|
| 61 |
+
to the value in `results.json` on the benchmark site. The `prompt_sha` field
|
| 62 |
+
lets you confirm the exact prompt template used.
|
| 63 |
+
|
| 64 |
+
## Links
|
| 65 |
+
|
| 66 |
+
- Benchmark site: [baseweight.co/benchmark](https://baseweight.co/benchmark)
|
| 67 |
+
- Pipeline code: [github.com/baseweight-ai/benchmark](https://github.com/baseweight-ai/benchmark)
|
| 68 |
+
- Prepared data splits: [baseweight-ai/baseweight-benchmark-data](https://huggingface.co/datasets/baseweight-ai/baseweight-benchmark-data)
|
| 69 |
+
- Fine-tuned adapters: [baseweight-ai/baseweight-benchmark-adapters](https://huggingface.co/collections/baseweight-ai/baseweight-benchmark-adapters-6a1613a5bee5213ab5e44936)
|