release-v1: LightGBM inference bundle + feature schema

README.md

@@ -0,0 +1,95 @@
+---
+language: en
+tags:
+  - twitter
+  - engagement-prediction
+  - lightgbm
+  - tabular
+license: apache-2.0
+library_name: lightgbm
+---
+
+# will-it-bang (release inference checkpoint)
+
+Predict a tweet's **exposure-adjusted engagement rate** from post-time text +
+caller-supplied ego/history features. This artifact is the shippable booster +
+transforms for the [will-it-bang](https://github.com/chriscarrollsmith/will-it-bang)
+research stack — **not** a fine-tuned language model.
+
+## Target
+
+`log1p(eng / (min(age_days, exposure_cap_days) + exposure_eps) ** exposure_beta)`
+
+where `eng = favorite_count + retweet_count` on the target tweet. Retweets
+(`full_text` starting with `RT @`) are excluded from training.
+
+## Metrics (time holdout, release knobs)
+
+| Metric | Value |
+|---|---:|
+| within-account Spearman NR (wmean) | 0.24558121396978394 |
+| within-account Spearman reply (wmean) | 0.3066146060439611 |
+| global Spearman NR | 0.6161772679216294 |
+| rate RMSE | 0.5877444773382154 |
+| features | 326 |
+
+Release train disables emb-kNN TE, CA provenance features, and domain-mean TE
+maps so serve-time does not need private train residue. `account_mean_rate` remains
+a **caller-supplied float** (recipe below).
+
+## Files
+
+| File | Role |
+|---|---|
+| `model_level.txt` | LightGBM level booster |
+| `model_rank_nr.txt` / `model_rank_rp.txt` | Optional hybrid stratum rankers |
+| `emb_pca.joblib` | MiniLM → PCA |
+| `tfidf.joblib` + `text_svd.joblib` | TF-IDF → TruncatedSVD |
+| `res_emb_cav_direction.npy` | Residual CAV direction |
+| `res_emb_baseline_coef.npy` + `res_emb_baselines.json` | OLS residualizer |
+| `feature_schema.json` | Ordered booster inputs + defaults |
+| `constants.json` | Priors / caps / encoder id |
+| `predict.py` | Reference scorer |
+
+## Input schema
+
+`account_id` is **not** a model input. Pass a feature row matching
+`feature_schema.json` (or raw `full_text` + Layer-A floats via `predict.py`).
+
+### `account_mean_rate` / `log_account_n`
+
+For a candidate at time T, take the author's earlier posts with known engagement,
+each converted to the same exposure-rate `y_i`. Let `n` be that count,
+`y_bar` their mean, `m = 20.0`, and
+`mu = 0.151304` (from `constants.json`):
+
+`account_mean_rate = (n * y_bar + m * mu) / (n + m)`  
+`log_account_n = log1p(n)`
+
+If `n = 0`: `account_mean_rate = mu`, `log_account_n = 0`.
+
+### Ego history
+
+Use pre-T reply/quote/mention arrivals and structural flags — **never** dump-time
+favorite/retweet history or current follower snapshots. See repo `AGENTS.md`.
+
+### Text encoder
+
+`sentence-transformers/all-MiniLM-L6-v2` (Apache-2.0), 384-d mean pool,
+then bundle residual-CAV → PCA (96 dims) and TF-IDF→SVD
+(128 dims).
+
+## Quick start
+
+```bash
+pip install lightgbm scikit-learn numpy pandas joblib pyarrow sentence-transformers
+python predict.py --bundle . --input row.json
+```
+
+`row.json` may include `full_text` plus any Layer-A fields from `feature_schema.json`.
+Missing fields use schema defaults.
+
+## Not included
+
+Training dataset, val predictions, per-account TE maps, emb-kNN train index,
+CA graph caches. Rebuild features yourself from the recipes above.

constants.json

@@ -0,0 +1,20 @@
+{
+  "global_mean": 0.15130374839379565,
+  "account_te_prior": 20.0,
+  "exposure_cap_days": 30.0,
+  "exposure_beta": 1.0,
+  "exposure_eps": 1.0,
+  "encoder_id": "sentence-transformers/all-MiniLM-L6-v2",
+  "encoder_dims": 384,
+  "emb_pca_dims": 96,
+  "text_svd_dims": 128,
+  "target": "log1p(eng / (min(age_days, exposure_cap_days) + exposure_eps) ** exposure_beta)",
+  "release_knobs": {
+    "account_te": true,
+    "emb_knn_te": false,
+    "domain_mean_te": false,
+    "provenance_features": false,
+    "residual_emb_cav": true,
+    "hybrid_rank": true
+  }
+}

emb_pca.joblib

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36ed38d2f2c853a29aa4c65761eac85207b610978c9f412147367083b61d2684
+size 151511

feature_schema.json

@@ -0,0 +1,2290 @@
+{
+  "features": [
+    {
+      "name": "text_len",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "log_text_len",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "has_url",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "n_mentions",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "n_hashtags",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "has_question",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "has_exclaim",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "n_emoji",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "starts_mention",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "caps_ratio",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "has_media",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "n_media",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "has_ca_url",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "n_ca_urls",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "is_reply",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "is_quote",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "is_self_reply",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "hour",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "dow",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "is_weekend",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "hour_sin",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "hour_cos",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "dow_sin",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "dow_cos",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "hist_n",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_tweets_7d",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_tweets_30d",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "days_since_last",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "hist_frac_reply",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_frac_url",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_frac_media",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_mean_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_median_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_p90_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_last_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_std_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_mean_replies_7d",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_mean_replies_30d",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_decay_mean_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_mean_replies_nonreply",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_mean_replies_reply",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "reply_x_hist_replies_nr",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "media_x_hist_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "structure"
+    },
+    {
+      "name": "hist_dump_n",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_dump_mean_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_dump_decay_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "hist_dump_p90_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "inbound_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "inbound_replies_7d",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "inbound_replies_30d",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "inbound_replies_per_tweet",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "inbound_quotes",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "inbound_quotes_7d",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "inbound_quotes_30d",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "inbound_quotes_per_tweet",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "inbound_mentions",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "inbound_mentions_7d",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "inbound_mentions_30d",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "inbound_mentions_per_tweet",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "media_x_inbound_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "reply_x_inbound_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "media_x_inbound_mentions",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "reply_x_inbound_mentions",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "ego_history"
+    },
+    {
+      "name": "has_reply_parent",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "reply_parent_emb_cos",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "reply_parent_log_text_len",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "reply_parent_is_reply",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "reply_parent_is_quote",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "reply_parent_same_author",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "reply_parent_age_hours",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "reply_parent_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "has_quote_parent",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "quote_parent_emb_cos",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "quote_parent_log_text_len",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "quote_parent_is_reply",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "quote_parent_is_quote",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "quote_parent_same_author",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "quote_parent_age_hours",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "quote_parent_replies",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "parent_context"
+    },
+    {
+      "name": "year",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "month_sin",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "month_cos",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "platform_era",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "era_pre_timeline",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "era_algo_timeline",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "era_pre_musk",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "era_musk_early",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "era_musk_late",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "calendar"
+    },
+    {
+      "name": "n_text_urls",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "url_structure"
+    },
+    {
+      "name": "url_only",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "url_structure"
+    },
+    {
+      "name": "url_at_end",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "url_structure"
+    },
+    {
+      "name": "has_expanded_domain",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "url_structure"
+    },
+    {
+      "name": "n_ca_url_rows",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "url_structure"
+    },
+    {
+      "name": "domain_is_twitter",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "url_structure"
+    },
+    {
+      "name": "domain_is_youtube",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "url_structure"
+    },
+    {
+      "name": "domain_is_github",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "url_structure"
+    },
+    {
+      "name": "domain_is_substack",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "url_structure"
+    },
+    {
+      "name": "domain_is_wiki",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "url_structure"
+    },
+    {
+      "name": "account_mean_rate",
+      "dtype": "float32",
+      "default": 0.15130374839379565,
+      "group": "account_prior"
+    },
+    {
+      "name": "log_account_n",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "account_prior"
+    },
+    {
+      "name": "res_emb_cav_proj",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "residual_cav"
+    },
+    {
+      "name": "res_emb_cav_proj_sq",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "residual_cav"
+    },
+    {
+      "name": "emb_0",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_1",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_2",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_3",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_4",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_5",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_6",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_7",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_8",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_9",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_10",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_11",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_12",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_13",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_14",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_15",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_16",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_17",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_18",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_19",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_20",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_21",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_22",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_23",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_24",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_25",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_26",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_27",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_28",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_29",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_30",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_31",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_32",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_33",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_34",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_35",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_36",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_37",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_38",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_39",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_40",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_41",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_42",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_43",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_44",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_45",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_46",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_47",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_48",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_49",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_50",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_51",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_52",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_53",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_54",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_55",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_56",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_57",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_58",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_59",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_60",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_61",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_62",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_63",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_64",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_65",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_66",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_67",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_68",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_69",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_70",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_71",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_72",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_73",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_74",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_75",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_76",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_77",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_78",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_79",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_80",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_81",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_82",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_83",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_84",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_85",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_86",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_87",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_88",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_89",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_90",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_91",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_92",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_93",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_94",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "emb_95",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "embedding_pca"
+    },
+    {
+      "name": "tfidf_0",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_1",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_2",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_3",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_4",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_5",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_6",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_7",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_8",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_9",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_10",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_11",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_12",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_13",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_14",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_15",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_16",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_17",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_18",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_19",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_20",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_21",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_22",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_23",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_24",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_25",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_26",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_27",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_28",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_29",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_30",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_31",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_32",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_33",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_34",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_35",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_36",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_37",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_38",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_39",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_40",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_41",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_42",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_43",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_44",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_45",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_46",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_47",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_48",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_49",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_50",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_51",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_52",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_53",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_54",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_55",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_56",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_57",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_58",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_59",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_60",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_61",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_62",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_63",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_64",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_65",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_66",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_67",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_68",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_69",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_70",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_71",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_72",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_73",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_74",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_75",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_76",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_77",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_78",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_79",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_80",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_81",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_82",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_83",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_84",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_85",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_86",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_87",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_88",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_89",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_90",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_91",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_92",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_93",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_94",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_95",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_96",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_97",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_98",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_99",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_100",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_101",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_102",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_103",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_104",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_105",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_106",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_107",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_108",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_109",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_110",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_111",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_112",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_113",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_114",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_115",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_116",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_117",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_118",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_119",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_120",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_121",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_122",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_123",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_124",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_125",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_126",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    },
+    {
+      "name": "tfidf_127",
+      "dtype": "float32",
+      "default": 0.0,
+      "group": "text_svd"
+    }
+  ],
+  "feature_names": [
+    "text_len",
+    "log_text_len",
+    "has_url",
+    "n_mentions",
+    "n_hashtags",
+    "has_question",
+    "has_exclaim",
+    "n_emoji",
+    "starts_mention",
+    "caps_ratio",
+    "has_media",
+    "n_media",
+    "has_ca_url",
+    "n_ca_urls",
+    "is_reply",
+    "is_quote",
+    "is_self_reply",
+    "hour",
+    "dow",
+    "is_weekend",
+    "hour_sin",
+    "hour_cos",
+    "dow_sin",
+    "dow_cos",
+    "hist_n",
+    "hist_tweets_7d",
+    "hist_tweets_30d",
+    "days_since_last",
+    "hist_frac_reply",
+    "hist_frac_url",
+    "hist_frac_media",
+    "hist_mean_replies",
+    "hist_median_replies",
+    "hist_p90_replies",
+    "hist_last_replies",
+    "hist_std_replies",
+    "hist_mean_replies_7d",
+    "hist_mean_replies_30d",
+    "hist_decay_mean_replies",
+    "hist_mean_replies_nonreply",
+    "hist_mean_replies_reply",
+    "reply_x_hist_replies_nr",
+    "media_x_hist_replies",
+    "hist_dump_n",
+    "hist_dump_mean_replies",
+    "hist_dump_decay_replies",
+    "hist_dump_p90_replies",
+    "inbound_replies",
+    "inbound_replies_7d",
+    "inbound_replies_30d",
+    "inbound_replies_per_tweet",
+    "inbound_quotes",
+    "inbound_quotes_7d",
+    "inbound_quotes_30d",
+    "inbound_quotes_per_tweet",
+    "inbound_mentions",
+    "inbound_mentions_7d",
+    "inbound_mentions_30d",
+    "inbound_mentions_per_tweet",
+    "media_x_inbound_replies",
+    "reply_x_inbound_replies",
+    "media_x_inbound_mentions",
+    "reply_x_inbound_mentions",
+    "has_reply_parent",
+    "reply_parent_emb_cos",
+    "reply_parent_log_text_len",
+    "reply_parent_is_reply",
+    "reply_parent_is_quote",
+    "reply_parent_same_author",
+    "reply_parent_age_hours",
+    "reply_parent_replies",
+    "has_quote_parent",
+    "quote_parent_emb_cos",
+    "quote_parent_log_text_len",
+    "quote_parent_is_reply",
+    "quote_parent_is_quote",
+    "quote_parent_same_author",
+    "quote_parent_age_hours",
+    "quote_parent_replies",
+    "year",
+    "month_sin",
+    "month_cos",
+    "platform_era",
+    "era_pre_timeline",
+    "era_algo_timeline",
+    "era_pre_musk",
+    "era_musk_early",
+    "era_musk_late",
+    "n_text_urls",
+    "url_only",
+    "url_at_end",
+    "has_expanded_domain",
+    "n_ca_url_rows",
+    "domain_is_twitter",
+    "domain_is_youtube",
+    "domain_is_github",
+    "domain_is_substack",
+    "domain_is_wiki",
+    "account_mean_rate",
+    "log_account_n",
+    "res_emb_cav_proj",
+    "res_emb_cav_proj_sq",
+    "emb_0",
+    "emb_1",
+    "emb_2",
+    "emb_3",
+    "emb_4",
+    "emb_5",
+    "emb_6",
+    "emb_7",
+    "emb_8",
+    "emb_9",
+    "emb_10",
+    "emb_11",
+    "emb_12",
+    "emb_13",
+    "emb_14",
+    "emb_15",
+    "emb_16",
+    "emb_17",
+    "emb_18",
+    "emb_19",
+    "emb_20",
+    "emb_21",
+    "emb_22",
+    "emb_23",
+    "emb_24",
+    "emb_25",
+    "emb_26",
+    "emb_27",
+    "emb_28",
+    "emb_29",
+    "emb_30",
+    "emb_31",
+    "emb_32",
+    "emb_33",
+    "emb_34",
+    "emb_35",
+    "emb_36",
+    "emb_37",
+    "emb_38",
+    "emb_39",
+    "emb_40",
+    "emb_41",
+    "emb_42",
+    "emb_43",
+    "emb_44",
+    "emb_45",
+    "emb_46",
+    "emb_47",
+    "emb_48",
+    "emb_49",
+    "emb_50",
+    "emb_51",
+    "emb_52",
+    "emb_53",
+    "emb_54",
+    "emb_55",
+    "emb_56",
+    "emb_57",
+    "emb_58",
+    "emb_59",
+    "emb_60",
+    "emb_61",
+    "emb_62",
+    "emb_63",
+    "emb_64",
+    "emb_65",
+    "emb_66",
+    "emb_67",
+    "emb_68",
+    "emb_69",
+    "emb_70",
+    "emb_71",
+    "emb_72",
+    "emb_73",
+    "emb_74",
+    "emb_75",
+    "emb_76",
+    "emb_77",
+    "emb_78",
+    "emb_79",
+    "emb_80",
+    "emb_81",
+    "emb_82",
+    "emb_83",
+    "emb_84",
+    "emb_85",
+    "emb_86",
+    "emb_87",
+    "emb_88",
+    "emb_89",
+    "emb_90",
+    "emb_91",
+    "emb_92",
+    "emb_93",
+    "emb_94",
+    "emb_95",
+    "tfidf_0",
+    "tfidf_1",
+    "tfidf_2",
+    "tfidf_3",
+    "tfidf_4",
+    "tfidf_5",
+    "tfidf_6",
+    "tfidf_7",
+    "tfidf_8",
+    "tfidf_9",
+    "tfidf_10",
+    "tfidf_11",
+    "tfidf_12",
+    "tfidf_13",
+    "tfidf_14",
+    "tfidf_15",
+    "tfidf_16",
+    "tfidf_17",
+    "tfidf_18",
+    "tfidf_19",
+    "tfidf_20",
+    "tfidf_21",
+    "tfidf_22",
+    "tfidf_23",
+    "tfidf_24",
+    "tfidf_25",
+    "tfidf_26",
+    "tfidf_27",
+    "tfidf_28",
+    "tfidf_29",
+    "tfidf_30",
+    "tfidf_31",
+    "tfidf_32",
+    "tfidf_33",
+    "tfidf_34",
+    "tfidf_35",
+    "tfidf_36",
+    "tfidf_37",
+    "tfidf_38",
+    "tfidf_39",
+    "tfidf_40",
+    "tfidf_41",
+    "tfidf_42",
+    "tfidf_43",
+    "tfidf_44",
+    "tfidf_45",
+    "tfidf_46",
+    "tfidf_47",
+    "tfidf_48",
+    "tfidf_49",
+    "tfidf_50",
+    "tfidf_51",
+    "tfidf_52",
+    "tfidf_53",
+    "tfidf_54",
+    "tfidf_55",
+    "tfidf_56",
+    "tfidf_57",
+    "tfidf_58",
+    "tfidf_59",
+    "tfidf_60",
+    "tfidf_61",
+    "tfidf_62",
+    "tfidf_63",
+    "tfidf_64",
+    "tfidf_65",
+    "tfidf_66",
+    "tfidf_67",
+    "tfidf_68",
+    "tfidf_69",
+    "tfidf_70",
+    "tfidf_71",
+    "tfidf_72",
+    "tfidf_73",
+    "tfidf_74",
+    "tfidf_75",
+    "tfidf_76",
+    "tfidf_77",
+    "tfidf_78",
+    "tfidf_79",
+    "tfidf_80",
+    "tfidf_81",
+    "tfidf_82",
+    "tfidf_83",
+    "tfidf_84",
+    "tfidf_85",
+    "tfidf_86",
+    "tfidf_87",
+    "tfidf_88",
+    "tfidf_89",
+    "tfidf_90",
+    "tfidf_91",
+    "tfidf_92",
+    "tfidf_93",
+    "tfidf_94",
+    "tfidf_95",
+    "tfidf_96",
+    "tfidf_97",
+    "tfidf_98",
+    "tfidf_99",
+    "tfidf_100",
+    "tfidf_101",
+    "tfidf_102",
+    "tfidf_103",
+    "tfidf_104",
+    "tfidf_105",
+    "tfidf_106",
+    "tfidf_107",
+    "tfidf_108",
+    "tfidf_109",
+    "tfidf_110",
+    "tfidf_111",
+    "tfidf_112",
+    "tfidf_113",
+    "tfidf_114",
+    "tfidf_115",
+    "tfidf_116",
+    "tfidf_117",
+    "tfidf_118",
+    "tfidf_119",
+    "tfidf_120",
+    "tfidf_121",
+    "tfidf_122",
+    "tfidf_123",
+    "tfidf_124",
+    "tfidf_125",
+    "tfidf_126",
+    "tfidf_127"
+  ],
+  "n_features": 326,
+  "notes": "Pass Layer-A floats (ego history, structure, calendar, account_mean_rate). Bundle applies residual CAV, emb PCA, and TF-IDF\u2192SVD. account_id is not a model input."
+}

metrics.json

@@ -0,0 +1,96 @@
+{
+  "rmse_log": 0.5877444773382154,
+  "mae_log": 0.2987727442764338,
+  "r2_log": 0.2190175711388792,
+  "spearman": 0.552975328379164,
+  "mae_engagement": 2.89875987639266,
+  "rmse_log_nonreply": 0.8638332178373588,
+  "rmse_log_reply": 0.3993264164625104,
+  "spearman_nonreply": 0.6161772679216294,
+  "spearman_reply": 0.4087985225427317,
+  "spearman_within_account_nonreply_mean": 0.2210206367964205,
+  "spearman_within_account_nonreply_wmean": 0.24558121396978394,
+  "spearman_within_account_nonreply_n": 415.0,
+  "spearman_within_account_reply_mean": 0.27944139633265186,
+  "spearman_within_account_reply_wmean": 0.3066146060439611,
+  "spearman_within_account_reply_n": 267.0,
+  "rmse_log_nonreply_top10": 2.3478436463235663,
+  "mean_underpred_nonreply_top10": 2.068597104093038,
+  "rmse_log_raw_dump": 1.0221516651925613,
+  "spearman_raw_dump": 0.448903523930651,
+  "exposure_correct": true,
+  "exposure_mode": "rate",
+  "exposure_cap_days": 30.0,
+  "exposure_beta": 1.0,
+  "exposure_a": 0.0,
+  "exposure_b": 0.0,
+  "target_space": "exposure_rate",
+  "best_iteration": 140,
+  "num_features": 326,
+  "train_rows": 88500,
+  "val_rows": 22125,
+  "n_accounts_train": 868,
+  "n_accounts_val": 594,
+  "baseline_rmse_log": 0.7146245709321322,
+  "reply_baseline_rmse_log": 0.7106109199414513,
+  "chosen_baseline_rmse_log": 0.6952656498247348,
+  "residual": false,
+  "baseline_kind": "context_replies",
+  "nonreply_weight": 1.0,
+  "breakout_head": false,
+  "breakout_auc": NaN,
+  "model": "lgbm",
+  "account_te": true,
+  "era_te": 0.0,
+  "era_te_prior": 0.0,
+  "era_heads": 0.0,
+  "era_key": "",
+  "era_head_min_train": 0.0,
+  "loss": "asymmetric",
+  "sera_floor": 0.0,
+  "sera_high_only": 0.0,
+  "sera_nonreply_only": 0.0,
+  "early_stop_metric": "rmse",
+  "hybrid_rank": 1.0,
+  "rank_nonreply_only": 1.0,
+  "rank_min_group": 2.0,
+  "rank_relevance": "dense",
+  "holdout_policy": "time",
+  "min_upload_hist": 0.0,
+  "provenance_features": 0.0,
+  "stream_fresh_days": 0.0,
+  "media_types": 0.0,
+  "nr_train_rank_n_rows": 14637.0,
+  "nr_train_rank_n_groups": 706.0,
+  "nr_train_rank_min_group": 2.0,
+  "nr_train_rank_mean_group": 20.73229461756374,
+  "nr_train_rank_frac_rows_kept": 0.9942263279445728,
+  "nr_train_rank_label_max": 174.0,
+  "nr_train_rank_relevance_extreme": 0.0,
+  "nr_train_rank_frac_grade4": 0.0,
+  "nr_val_rank_n_rows": 6891.0,
+  "nr_val_rank_n_groups": 420.0,
+  "nr_val_rank_min_group": 2.0,
+  "nr_val_rank_mean_group": 16.40714285714286,
+  "nr_val_rank_frac_rows_kept": 0.9826037359190075,
+  "nr_val_rank_label_max": 127.0,
+  "nr_val_rank_relevance_extreme": 0.0,
+  "nr_val_rank_frac_grade4": 0.0,
+  "reply_train_rank_n_rows": 73776.0,
+  "reply_train_rank_n_groups": 865.0,
+  "reply_train_rank_min_group": 2.0,
+  "reply_train_rank_mean_group": 85.29017341040462,
+  "reply_train_rank_frac_rows_kept": 0.9999728916479167,
+  "reply_train_rank_label_max": 174.0,
+  "reply_train_rank_relevance_extreme": 0.0,
+  "reply_train_rank_frac_grade4": 0.0,
+  "reply_val_rank_n_rows": 14994.0,
+  "reply_val_rank_n_groups": 268.0,
+  "reply_val_rank_min_group": 2.0,
+  "reply_val_rank_mean_group": 55.94776119402985,
+  "reply_val_rank_frac_rows_kept": 0.9921916357861302,
+  "reply_val_rank_label_max": 170.0,
+  "reply_val_rank_relevance_extreme": 0.0,
+  "reply_val_rank_frac_grade4": 0.0,
+  "seconds": 26.48
+}

predict.py

@@ -0,0 +1,173 @@
+"""Reference scorer for a will-it-bang release inference bundle.
+
+Usage:
+    uv run predict.py --bundle ~/.cache/will-it-bang/release-v1 --input row.json
+    uv run predict.py --bundle DIR --input rows.jsonl
+
+Input JSON object(s) may include `full_text` plus any Layer-A fields from
+`feature_schema.json`. Missing fields use schema defaults. MiniLM is computed
+from `full_text` when raw `emb_0..383` are absent.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import re
+from pathlib import Path
+from typing import Any
+
+import joblib
+import lightgbm as lgb
+import numpy as np
+import pandas as pd
+
+
+def _load_json_rows(path: Path) -> list[dict[str, Any]]:
+    text = path.read_text()
+    if path.suffix == ".jsonl":
+        return [json.loads(line) for line in text.splitlines() if line.strip()]
+    payload = json.loads(text)
+    if isinstance(payload, list):
+        return payload
+    if isinstance(payload, dict):
+        return [payload]
+    raise SystemExit(f"Unsupported JSON in {path}")
+
+
+def _minilm_embed(texts: list[str], encoder_id: str) -> np.ndarray:
+    from sentence_transformers import SentenceTransformer
+
+    model = SentenceTransformer(encoder_id)
+    emb = model.encode(texts, show_progress_bar=False, normalize_embeddings=False)
+    return np.asarray(emb, dtype=np.float64)
+
+
+def _raw_emb_matrix(rows: list[dict[str, Any]], n_dim: int) -> np.ndarray | None:
+    cols = [f"emb_{i}" for i in range(n_dim)]
+    if not all(c in rows[0] for c in cols):
+        return None
+    return np.asarray([[float(r.get(c, 0.0)) for c in cols] for r in rows], dtype=np.float64)
+
+
+def build_feature_matrix(
+    rows: list[dict[str, Any]],
+    *,
+    bundle: Path,
+) -> tuple[np.ndarray, list[str]]:
+    schema = json.loads((bundle / "feature_schema.json").read_text())
+    constants = json.loads((bundle / "constants.json").read_text())
+    feature_names: list[str] = list(schema["feature_names"])
+    defaults = {f["name"]: float(f["default"]) for f in schema["features"]}
+
+    pca = joblib.load(bundle / "emb_pca.joblib")
+    tfidf = joblib.load(bundle / "tfidf.joblib")
+    svd = joblib.load(bundle / "text_svd.joblib")
+    direction = np.load(bundle / "res_emb_cav_direction.npy")
+    coef = np.load(bundle / "res_emb_baseline_coef.npy")
+    baselines = json.loads((bundle / "res_emb_baselines.json").read_text())
+    baseline_names: list[str] = list(baselines["names"])
+
+    n_raw = int(constants["encoder_dims"])
+    emb = _raw_emb_matrix(rows, n_raw)
+    if emb is None:
+        texts = [str(r.get("full_text", "")) for r in rows]
+        if not any(texts):
+            raise SystemExit("Need full_text or emb_0..emb_{n-1} on each row")
+        emb = _minilm_embed(texts, constants["encoder_id"])
+        if emb.shape[1] != n_raw:
+            raise SystemExit(f"Encoder dim {emb.shape[1]} != {n_raw}")
+
+    # Layer-A frame for residual baselines + booster columns.
+    frame = pd.DataFrame(rows)
+    for name, default in defaults.items():
+        if name not in frame.columns and not re.fullmatch(r"emb_\d+", name) and not name.startswith(
+            "tfidf_"
+        ):
+            frame[name] = default
+        elif name in frame.columns:
+            frame[name] = pd.to_numeric(frame[name], errors="coerce").fillna(default)
+
+    # Residualize raw MiniLM with shipped OLS coeffs, then CAV projection.
+    z_cols: list[np.ndarray] = []
+    for name in baseline_names:
+        if name == "intercept":
+            z_cols.append(np.ones(len(rows), dtype=np.float64))
+            continue
+        col = frame[name].to_numpy(dtype=np.float64) if name in frame.columns else np.zeros(len(rows))
+        z_cols.append(np.nan_to_num(col, nan=0.0, posinf=0.0, neginf=0.0))
+    z = np.column_stack(z_cols)
+    if z.shape[1] != coef.shape[0]:
+        raise SystemExit(
+            f"Residual baseline width {z.shape[1]} != coef rows {coef.shape[0]} "
+            f"(names={baseline_names})"
+        )
+    resid = emb - z @ coef
+    proj = resid @ direction.reshape(-1)
+    frame["res_emb_cav_proj"] = proj.astype(np.float32)
+    frame["res_emb_cav_proj_sq"] = (proj * proj).astype(np.float32)
+
+    emb_pca = pca.transform(emb.astype(np.float32))
+    emb_pca_df = pd.DataFrame(
+        emb_pca, columns=[f"emb_{i}" for i in range(emb_pca.shape[1])]
+    )
+    texts = [str(r.get("full_text", "")) for r in rows]
+    tfidf_z = svd.transform(tfidf.transform(texts)).astype(np.float32)
+    tfidf_df = pd.DataFrame(
+        tfidf_z, columns=[f"tfidf_{i}" for i in range(tfidf_z.shape[1])]
+    )
+    drop_emb = [c for c in frame.columns if re.fullmatch(r"emb_\d+", c)]
+    if drop_emb:
+        frame = frame.drop(columns=drop_emb)
+    frame = pd.concat(
+        [frame.reset_index(drop=True), emb_pca_df, tfidf_df],
+        axis=1,
+    )
+
+    x = np.column_stack(
+        [
+            pd.to_numeric(frame[c], errors="coerce").fillna(defaults.get(c, 0.0)).to_numpy(
+                dtype=np.float32
+            )
+            if c in frame.columns
+            else np.full(len(rows), defaults.get(c, 0.0), dtype=np.float32)
+            for c in feature_names
+        ]
+    )
+    return x, feature_names
+
+
+def predict_rows(bundle: Path, rows: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    x, _names = build_feature_matrix(rows, bundle=bundle)
+    model = lgb.Booster(model_file=str(bundle / "model_level.txt"))
+    scores = np.asarray(model.predict(x), dtype=np.float64)
+    out: list[dict[str, Any]] = []
+    for row, score in zip(rows, scores, strict=True):
+        item = {
+            "y_pred": float(score),
+            "y_pred_eng_rate": float(np.expm1(np.clip(score, -1.0, 30.0))),
+        }
+        if "tweet_id" in row:
+            item["tweet_id"] = row["tweet_id"]
+        out.append(item)
+    return out
+
+
+def main() -> None:
+    p = argparse.ArgumentParser(description=__doc__)
+    p.add_argument("--bundle", type=Path, required=True, help="Release bundle directory")
+    p.add_argument("--input", type=Path, required=True, help="JSON or JSONL feature rows")
+    p.add_argument("--output", type=Path, default=None, help="Write predictions JSON")
+    args = p.parse_args()
+    rows = _load_json_rows(args.input)
+    preds = predict_rows(args.bundle, rows)
+    text = json.dumps(preds, indent=2) + "\n"
+    if args.output:
+        args.output.write_text(text)
+        print(f"Wrote {len(preds)} predictions → {args.output}")
+    else:
+        print(text, end="")
+
+
+if __name__ == "__main__":
+    main()

res_emb_baseline_coef.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:393300a86a51f33f40905ac91a7983457a8b7ff2efcf5f2c59ad79a79462e39d
+size 12416

res_emb_baselines.json

@@ -0,0 +1,10 @@
+{
+  "names": [
+    "intercept",
+    "account_mean_rate",
+    "is_reply",
+    "inbound_replies_per_tweet"
+  ],
+  "quantile": 0.1,
+  "nonreply_only": true
+}

res_emb_cav_direction.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8d1660dc66d9c11429345a0efcc14f4edb204b13c2785388d03bcf4f4dc4b12
+size 3200

text_svd.joblib

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:32132588125026dcadfc3c12d917fafa44d5a0c694642453e461066c1261ea29
+size 51203815

tfidf.joblib

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a9878d4c03ecbdf0a46f85dccdf53380168e9c93fd2e6cfcb406cb27e31d0a7
+size 1962598