philipstevens's picture
Add dataset card
5992e42 verified
|
Raw
History Blame Contribute Delete
2.73 kB
metadata
license: apache-2.0
tags:
  - baseweight-benchmark
  - benchmark
  - predictions

Baseweight Benchmark — Raw Predictions

Every model generation produced during evaluation for the Baseweight Benchmark, a reproducible comparison of fine-tuned open models against frontier APIs on focused vertical tasks.

What's here

One JSONL file per model × task × condition × seed. Each line is one prediction:

Field Description
id Row identifier from the prepared test split
model Model short name (qwen3-8b, gpt-5.4-mini)
condition lora, zero-shot, or 5-shot
eval_seed Seed used for this eval run (0, 1, or 2)
prompt_sha SHA-256 prefix of the prompt template used
input The text input sent to the model
output The raw model output
ground_truth Gold label or answer
input_tokens Prompt token count
output_tokens Completion token count
avg_logprob Mean log-probability of output tokens
timestamp UTC time of the generation

File layout

predictions/
  local/
    qwen3-8b/
      banking77/   lora.jsonl  lora_seed1.jsonl  lora_seed2.jsonl
                   zero-shot.jsonl  zero-shot_seed1.jsonl  zero-shot_seed2.jsonl
                   5-shot.jsonl  5-shot_seed1.jsonl  5-shot_seed2.jsonl
      cuad/        lora.jsonl  lora_seed1.jsonl  lora_seed2.jsonl
                   zero-shot.jsonl  zero-shot_seed1.jsonl  zero-shot_seed2.jsonl
  api/
    gpt-5.4-mini/
      banking77/   zero-shot.jsonl  zero-shot_seed1.jsonl  zero-shot_seed2.jsonl
                   5-shot.jsonl  5-shot_seed1.jsonl  5-shot_seed2.jsonl
      cuad/        zero-shot.jsonl  zero-shot_seed1.jsonl  zero-shot_seed2.jsonl

Seeds 1 and 2 are independent re-evaluations with shuffled test order. Seed 0 is the base run (file without seed suffix). Reported metrics are the mean across all three seeds; 95% CIs are bootstrap intervals over the seeds.

Verifying the numbers

Download any JSONL, recompute the metric against ground_truth, and compare to the value in results.json on the benchmark site. The prompt_sha field lets you confirm the exact prompt template used.

Links