---
license: apache-2.0
tags: [nla, qwen3, length-penalty, co-training]
configs:
  - config_name: heldout
    data_files: heldout_ct/*.samples.jsonl
---

# NLA co-train comparison — held-out samples

Held-out completions (n=1000, rows 25k:26k, doc-disjoint from RL train) for the co-trained λ=0.006 / λ=0.015 models, each scored two ways:

- `*_ct_baseAR` — scored by the frozen base AR (same ruler as the frozen sweep).
- `*_ct_ownAR` — scored by the model's own co-trained AR (real co-trained system FVE).

Per-sample rows: `idx, n_tokens, fve, nmse, mse, reward, extracted, cjk, explanation, source_text`. Matched by `idx` across models and against the frozen sweep's `-results` dataset. `COTRAIN_COMPARISON.md` has the headline table.