GunGG4/paiec-Qwen8Binstruct-guneesh-data

PAIEC: Qwen3-VL-8B-Instruct item embeddings and IRT fits

Training artifacts made for the Stanford AIMS Predictive AI Evaluation Competition (PAIEC), NeurIPS 2026. Everything here is derived from the public training data in aims-foundations/measurement-db. No test data is included.

Code that made these files and uses them: branch Guneesh of https://github.com/jaipack17/AIMS-Competition, folder guneesh/ (see its README.md, sections 4 and 9).

Download

From inside the guneesh/ folder of the code:

hf download GunGG4/paiec-Qwen8Binstruct-guneesh-data --repo-type dataset --local-dir data

The files then sit at the paths the scripts expect (data/emb/..., data/irt/...).

Contents (about 13.6 GB: 2.5 GB of embeddings and 1D IRT, 11 GB of MIRT fits)

emb/qwen3vl8b/ — item embeddings (1.9 GB)

• One .npz per binary training benchmark (115 files), plus _progress.csv (items, unique texts, tokens and seconds per benchmark).

• 255,286 items: up to 5,000 per benchmark, drawn at random (seeded by crc32 of the benchmark name) from the items with at least one response. 33 benchmarks were capped at 5,000; the other 82 are complete.

• Model: Qwen/Qwen3-VL-8B-Instruct, bfloat16. Vector = last layer, last token hidden state, 4096 numbers (float32). Max 8192 tokens, truncation from the left, batch size 1, seed 0, deterministic.

• Text: the Agent Psychometrics "no solution" prompt:

  Task statement:
  {item content}
  
  How difficult is the above task for a coding agent? Please output one floating-point number from 0 (very easy) to 1 (very hard). Your difficulty score:
  

• Keys in each file (the Agent Psychometrics cache format): task_ids (item id inside the benchmark; the full id elsewhere is <benchmark>::<task_id>), X (n × 4096 float32), counts (text length in characters), and metadata (backbone, max_length, embedding_dim, embedding_layer = −1, instruction, text_format, ...).

irt/ — response tables

File	Rows	Columns
cells_binary.parquet	24,998,450	pair, item, n (attempts), k (correct), benchmark
cells_cont.parquet	3,553,143	pair, item, n, sy (sum of y scaled to [0,1]), sz, sz2 (sum and sum of squares of logit(clip(y, 0.01, 0.99))), benchmark
subjects.parquet	13,658	the 8 subject fields, subject_id, display_name, benchmark, pair, and the normalised parts name, harness_n, provider_n, effort_n
folds.csv	115	benchmark, fold (0–4): the 115 binary benchmarks split at random (seed 0) into 5 folds of 23

A cell is one subject on one item, with repeated attempts kept as counts. pair = <benchmark>::<subject_id>, item = <benchmark>::<item_id>. 152 training benchmarks: 115 binary and 37 non-binary.

irt/versionB_4parts_logitnormal/ — IRT fits (300 MB)

Six fits: fold0 … fold4 (each leaves out that fold's 23 binary benchmarks entirely) and final (all data). Model (1PL, fitted with Pyro SVI as in Agent Psychometrics: Trace_ELBO, ClippedAdam lr 0.01, 2000 epochs, full batch):

mu = theta_name + theta_harness + theta_provider + theta_effort + zeta(pair) - b(item)
binary:      outcome ~ Bernoulli(sigmoid(mu))
non-binary:  logit(y) ~ Normal(mu, sigma_m^2), one sigma_m per benchmark

File	Columns
items.parquet	item, b (difficulty, posterior mean), b_scale (posterior std)
theta_name/harness/provider/effort.parquet	key, theta, scale
zeta.parquet	pair, zeta, scale
hyper.json	learned prior spreads (final: sigma_name 0.377, sigma_harness 0.647, sigma_provider 0.455, sigma_effort 0.528, sigma_zeta 1.110, sigma_b 1.811), per-benchmark sigma_m, settings
losses.csv	training loss per epoch
name_date_prior.json	release-date line for unseen model names (theta_name = a + b · years since 2024)

Names that were blank in the source get a private key private::<benchmark>::<subject_id>.

irt/mirt_<form>_K<K>/fold<f>/ — multidimensional IRT fits (11 GB, 70 fits)

70 multidimensional IRT (MIRT) fits on the same data and folds as the 1D fits: forms nested (theta + zeta + u·v − b, the 1D model plus K extra dimensions) and m2pl (a·theta − b, the standard compensatory multidimensional 2PL), K = 1, 2, 3, 5, 8, 10, 15, folds 0–4. Columns, equations, fitting settings, a usage example, the AIC/BIC table (irt/mirt_fit_summary.csv) and what we found when using them for new benchmarks are in irt/MIRT_README.md. The code is in code/ (mirt_fit.py, modal_mirt.py).

Download only these: hf download GunGG4/paiec-Qwen8Binstruct-guneesh-data --repo-type dataset --local-dir data --include "irt/mirt_*"

Licence and credit

These files are derived from aims-foundations/measurement-db, which is released under CC-BY-SA-4.0, so they are shared under the same licence. The original upstream licences and limitations of each benchmark still apply. Please credit aims-foundations/measurement-db and the original benchmark authors.

Embeddings were made with Qwen/Qwen3-VL-8B-Instruct (Apache 2.0) and the embedding code of Agent Psychometrics (Ge, Kryvosheieva, Fried, Girit, Hariharan, COLM 2026; MIT licence).