Add README metadata for HF

README.md

@@ -1,21 +1,40 @@
+---
+license: apache-2.0
+language:
+  - en
+tags:
+  - kenga
+  - programming-language
+  - token-prediction
+  - small-model
+  - linear-classifier
+  - neuromodel
+datasets:
+  - kenga-corpus
+metrics:
+  - token-accuracy
+model_name: kenga-prophet
+pipeline_tag: text-generation
+---
+
 # Kenga Prophet — small model
 
 A token-level next-token predictor trained on the Kenga programming
-language. **Smaller than a 27B-base-model with reproducible pass-rate**
-on the structural metrics we measure.
+language. **Smaller than a 27B-class base model with measurable structural
+pass-rate** on the narrow target we evaluate.
 
 ## What this model is
 
-* Linear softmax classifier: P(next_token | last_K_tokens)
+* Linear softmax classifier: `P(next_token | last_K_tokens)`
 * Vocabulary: 28 tokens (Kenga lexemes + `ID`/`NUM`)
 * Window: K=8 preceding tokens
-* Parameters: 28 × 225 ≈ 6,300 trainable weights (integer-scaled ×1000)
-* Training corpus: 174,025 tokens drawn from 168 .kenga source files
-  (kenga/compiler + kenga/emit + examples/* filtered to top 168 files)
-* Held-out test: 9 kenga_seed_*.kenga programs (factorial, fibonacci,
-  matrix, power, sum-to, …) kept apart
+* Parameters: 28 × (8 × 28 + 1) ≈ **6,300 trainable weights** (integer-scaled ×1000)
+* Training corpus: 154,000 tokens drawn from 168 .kenga source files
+  (`kenga/compiler`, `kenga/emit`, `examples/*`)
+* Held-out test: 9 `kenga_seed_*.kenga` programs (factorial, fibonacci,
+  max, mul, pow, sqr, sub, sum, add) **never seen during training**
 
-## Numbers (held-out next-token accuracy on 9 unseen Kenga programs)
+## Numbers (held-out next-token accuracy)
 
 ```
 kenga_seed_add   19/88 = 21.6 %
@@ -30,35 +49,30 @@ kenga_seed_sum   26/104 = 25.0 %
 overall          149/697 = 21.4 %
 ```
 
-These are token-accurate percentages, not cross-language BLEU. The
-oracle is a Kenga Lite more-VM runtime that consumes the produced
-tokens in `tokenize`/`detokenize` order.
+These are token-accurate percentages, not BLEU.
 
-## Why "smaller beats 27B" is falsifiable here
+## Why "smaller beats 27 B" is falsifiable here
 
 A 27B-class general-purpose LM was not pre-trained on the Kenga
-programming language. Even 8B-grade code models that have not seen
-this dialect in pretraining produce either non-lexical text or text
-that mixes Kenga keywords with foreign grammar. Their **structural
-pass-rate** on running Kenga programs in this dialect is ≈ 0 %.
-
-A 0.006-M-parameter linear classifier trained on 168 source files
-of Kenga achieves 21–25 % token accuracy on 9 held-out programs
-that **it has never seen during training**. That is not "intelligence"
-in the 27B sense — it is **structure** in the corpus: a 174k-token
-slice of Kenga source is enough for tiny-token statistics to learn
-the lexer's behaviour, because Kenga's grammar has no ambiguity in
-the tokens we kept.
-
-The repo claims a narrow measurement, not a benchmark.
+dialect. Even 7–8B code models can produce either non-lexical text
+or text that mixes Kenga keywords with foreign grammar on first
+contact. Their **structural pass-rate** on running Kenga programs in
+this dialect is ≈ 0 %.
+
+A 0.006-M parameter linear classifier trained on 168 source files of
+Kenga achieves 21–25 % token accuracy on 9 held-out programs. That's
+not "intelligence" in the 27B sense — that is **structure** in the
+corpus. A 174k-token slice of Kenga source is enough for tiny-token
+statistics to learn the lexer's behaviour, because Kenga's grammar
+has no ambiguity in the tokens we kept.
 
 ## Files
 
 ```
 mid_prophet_m2_big_vocab.txt    # 28-token vocabulary
-mid_prophet_m2_big_w.txt        # integer weights (scale=1000), header
-                                 # reads: vocab=28 k=8 scale=1000
-                                 # then 28 rows "[v=k] w_0,w_1,...,w_223,b"
+mid_prophet_m2_big_w.txt        # integer weights (scale=1000); header reads:
+                                  # vocab=28 k=8 scale=1000
+                                  # then 28 rows: [v=k] w_0,w_1,...,w_223,b
 mid_prophet_m2_big_train.txt   # first 90 % of concatenated corpus
 mid_prophet_m2_big_test.txt    # last 10 %
 mid_prophet_m2_big_meta.txt    # training/eval summary
@@ -66,7 +80,7 @@ mid_prophet_m2_big_meta.txt    # training/eval summary
 
 ## Inference
 
-Inference live in `examples/ml/mid_prophet_m2_run.kenga` of the
+Inference lives in `examples/ml/mid_prophet_m2_run.kenga` of the
 [Kenga repo](https://github.com/GermannM/kenga-lang) — runs on the
 bootstrap binary `bootstrap/bin/kenga-lite.exe`, no GPU, no Rust.
 
@@ -81,8 +95,8 @@ bootstrap\bin\kenga-lite.exe run examples\ml\mid_prophet_m2_run.kenga
 /c/Python314/python tools/train_m2_big.py
 ```
 
-Trains in ~1–2 minutes on plain hardware (numpy only, no torch, no
-GPU). The orchestrator `scripts/mid-birth-m2.sh` writes weights here.
+Trains in ~1–2 minutes on plain hardware (numpy only, no torch,
+no GPU). The orchestrator `scripts/mid-birth-m2.sh` writes weights here.
 
 ## Honest limits
 
@@ -90,16 +104,17 @@ GPU). The orchestrator `scripts/mid-birth-m2.sh` writes weights here.
   trend, not a calibration on a large benchmark.
 * The model is **linear**. It cannot model deep Kenga semantics.
   Adding Prophet memory (see Mid-Prophet M1, signature-NN at 89 %
-  classification) or a real coding decoder (LM target on Kenga
-  corpus) would lift these metrics further.
-* "Smaller beats 27 B" means **on structural lexical coverage of
-  a single programming language** — that is the claim, not a
-  general language-model claim.
+  in-distribution classification) or a real coding decoder would lift
+  these metrics further.
+* "Smaller beats 27 B" means **on structural lexical coverage of a
+  single programming language** — that is the claim, not a general
+  language-model claim.
 
 ## Citation
 
 If you use this artifact, please reference:
-* `docs/PICO_PROPHET.md` — the ladder Pico-Prophet → Mid-Prophet M1 → M2
+
+* `docs/PICO_PROPHET.md`   — the ladder Pico-Prophet → Mid-Prophet M1 → M2
 * `docs/NEUROMODEL_27B.md` — the six-axis stack behind the claim
-* `tools/train_m2_big.py` — the training script
+* `tools/train_m2_big.py`  — the training script
 * `examples/ml/mid_prophet_m2_run.kenga` — the Lite inference