File size: 1,415 Bytes
a52b535 | 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 | ---
license: cc-by-4.0
language: [en]
tags: [psychometrics, quantization, IRT, DIF, LLM-evaluation]
---
# quant-dif-raw v1
Raw response data for **"Quantization as a Measurement Invariance Problem: A
Differential Item Functioning Analysis of Psychometric Scales in Compressed
LLMs"**.
- 189,000 logged administrations of the 50-item IPIP Big-Five Factor Markers
to 7 frozen model configurations (Qwen2.5-3B fp16/q8_0/q4_0, Qwen2.5-7B
q8_0/q4_0, Llama-3.2-3B q8_0/q4_0), served by Ollama v0.31.1.
- Conditions: main experiment (200 personas × 50 items), no-persona
distribution sub-experiment (40 samples/item), paraphrase-bank rerun,
temperature 0.2/1.0 subsets.
- Every JSONL record carries full provenance: model tag, precision label,
persona/item ids, per-call seed, temperature, GGUF weight SHA-256, engine
version, timestamp, raw response text, parsed 1–5 score, and top-5 token
log-probabilities.
- Bundle also contains the human-verified paraphrase item bank, the 200-persona
pool (deterministic, seed 42), and the model manifest.
`quant-dif-raw-v1.tar.gz` SHA-256:
`5b03248483fb60871a802331614e11af9b4ecb2c44321e1c9e1ccf580f36ef96`
Analysis code and preregistration: https://github.com/XuebinMa/quant-dif
(preregistration frozen at commit `a1a8582` before any data collection).
Note: IPIP item texts (public domain, ipip.ori.org) appear inside prompt/record
fields as administered.
|