box C pool_b: 1,247 rows, 20 languages, orthography scored

270m_pool_b/README.md

@@ -0,0 +1,25 @@
+# Box C generated text — `pool_b`
+
+1,247 rows across 20 languages, generated on CPU with `google/gemma-3-270m`
+and scored with this project's rule-based metrics, **including `tier2_orthography`**
+(diacritic density against each language's own corpus rate — the strongest single
+discriminator in the rule set, AUC 0.955 separating generated from real text).
+
+## Prompt system: `pool_b`
+
+English instruction prompts crossing 20 domains × 14 genres × 3 lengths × 4 registers. **These did not work as intended**: a base checkpoint continues an instruction block rather than obeying it, so ~99% of these rows are English boilerplate (`"Do not use quotation marks, bullet points or headings"`) carrying an African language code.
+
+Published anyway because wrong-language output is a real floor class for a quality estimator — but never pool it with in-language rows, or the model learns that Yoruba looks like English.
+
+## Columns
+
+`prompt_system` and `seed_source` identify the generation regime; `source_box`
+records the machine. Every other column matches the main dataset's schema.
+
+## Honest limits
+
+- `gemma-3-270m` is a base checkpoint, not instruction-tuned, and has no chat
+  template. It continues text; it does not follow instructions.
+- Rows where `tier1_is_text` is false have no `tier2_*` scores: the gate short-circuits
+  grading. They are still informative — they failed on digit runs and repetition,
+  which is exactly the degenerate behaviour worth labelling.

270m_pool_b/data.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41f823c2385249541eeb87785836bc29730caa3d9dc3c2b89a160fdeb2639cab
+size 885257