README.md

---
license: cc-by-4.0
language: en
tags:
  - babylm
  - babylm-2026
  - strict-small
  - evaluation-artifacts
  - reproducibility
pretty_name: bind1 (BabyLM 2026 Strict-Small) — full evaluation artifacts
---

# bind1 — BabyLM 2026 Strict-Small entry: full evaluation artifacts

Complete, per-item evaluation outputs for the
[`SecludedCorner/bind1-babylm2026-strict-small`](https://huggingface.co/SecludedCorner/bind1-babylm2026-strict-small)
entry and its ablation families, released so that **every number we report is downloadable and
recomputable** — including the negative results.

All outputs were produced by the official `babylm-eval` pipeline (zero-shot backend `causal`;
GLUE via the official fine-tuning protocol). Reports are the pipeline's own
`best_temperature_report.txt`; predictions are per-item.

## Model families included (35 evaluation targets)

| Directory pattern | What it is | What it shows |
|---|---|---|
| `hf_loop2_full(_sN)` | the entry architecture, trained loop depth T=3, seeds 0–3 | main results |
| `hf_loop2_T1/T2/T4(_sN)` | same weights, forced inference depth T=1/2/4 | the T-sweep: iteration helps grammar in all seeds, erodes healthy-write entity tracking, and partially rebuilds the one collapsed write |
| `hf_loop2_sever(_sN)` | same weights, verdict-to-trust edge zeroed at inference | inference-inertness of the feedback edge (4/4 seeds) |
| `hf_loop2_novg(_sN)` (+`_ti1`) | retrained from scratch with the edge frozen | training-scaffold evidence (seed-conditional: 3/4) |
| `hf_baseline_full(_sN)` | size-matched plain-decoder monolith, seeds 0–3 | the paired control for every claim |
| `hf_loop_full` | the earlier bypassed variant (v=1.0, loop inert) | the audit-caught failure kept as a diagnostic |

Internal (logical) experiment ids map to these physical names as
`bind1_tt3_sN` = `hf_loop2_full(_sN)`, `bind1_tt3_sN_ti{K}` = `hf_loop2_T{K}(_sN)`,
`mono_sN` = `hf_baseline_full(_sN)`, `bind1_tt3_sN_novg` = `hf_loop2_novg(_sN)`.

## Layout

```
results/<model>/zero_shot/causal/<task>/<subset>/
    best_temperature_report.txt   # official aggregate report
    predictions.json              # per-item predictions
results/hf_loop2_full/finetune/<task>/results.txt      # official GLUE protocol (entry)
results/hf_baseline_full/finetune/<task>/results.txt   # official GLUE protocol (control)
scripts/oracle_best_of_T.py       # best-of-T oracle + chance-null recovery analysis
scripts/make_t_variant.py         # builds the forced-T inference variants (config-only clones)
patches/compute_results_per_item_is_correct.diff  # additive patch to the official pipeline
```

## Per-item correctness (`is_correct`)

Entity-tracking `predictions.json` records include `pred_idx` / `gold_idx` / `is_correct`,
added by the **additive** patch in `patches/` (the official `id`/`pred` fields and all scoring
logic are untouched; patched reruns reproduce the official aggregates exactly, 12/12 checked).
This is what makes the chance-null analysis possible: for answers the single-pass model got
wrong and iteration *changed*, a memoryless re-roll lands on gold ~25% of the time (5 options).
Recovery at ~25% = re-rolled noise; recovery far above = signal. Both verdicts occur in
this data **under the pre-revision item set** — ~25% in the three healthy-write seeds, 37–45% in the
collapsed-write seed — which was the evidence behind the model card's original "iteration as a backup
pathway" note.

**Post-revision update (2026-07-09):** under the leaderboard's revised standard (drop "nothing"-gold
items) the 37–45% cells fall to 23.4–24.7% ≈ null; the model card note has been updated accordingly and
the backup-pathway reading retired. The per-item files here let you reproduce both readings.

Recompute it yourself:

```bash
python scripts/oracle_best_of_T.py bind1_tt3_s0   # any of s0..s3
```

(The script expects the experiment-archive layout; adapt the paths at the top, or read the
predictions files directly — the schema is one JSON object per subset with a `predictions` list.)

## Scope note

This dataset now covers **both** the entry generation's ablation families and the
**training-time-depth study** (`hf_bind1_tt1_s0..s9`: ten seeds trained single-pass — the
capability never forms; see the ablations model repo for the loadable checkpoints). Nothing in this dataset
was selected post-hoc: every ablation family we ran on the entry generation is present,
including the ones that falsified our own pre-registered predictions.

## Citation

Yulin Yang (ORCID 0009-0007-4827-8449). *A Microkernel Language Model: Reasoning as
Mutually-Supporting Aggregates, and Why Its Parts Must Be Judged Together.* BabyLM Challenge
2026 (Strict-Small track) submission.