Add benchmark tables (external BPB, held-out BPB, TatBLiMP)

README.md

@@ -38,9 +38,49 @@ The tokenizer is SentencePiece-based; pass `use_fast=False` so tokenization matc
 Two stages. Stage 1 pretrains on a Cyrillic-Qypchaq pool (Tatar 43 / Kazakh 30 / Kyrgyz 16 / Bashkir 10 / smaller siblings ~1 percent) under target-capped sampling. Stage 2 specializes on pure, register-weighted literary Tatar with an anti-memorization regime that keeps verbatim reproduction low.
 
 ## Results
-- TatBLiMP (grammaticality): acc 0.975 / acc_norm 0.958, tying a monolingual 125M baseline and ahead of a 7B adaptation (0.956) and a 1.3B one (0.736). Four frontier models from 31B to 120B all score below TATlit and below the 125M baseline.
-- Byte-per-byte compression: leads every held-out and external set, widest on the literary register, ahead of all frontier models up to 120B.
-- Knowledge (TUMLU-mini): near chance by design.
+
+TATlit is compared against three Tatar-specialized baselines (Goldfish-tat 125M, Tweety-7B-tatar, mGPT-1.3B-tatar) and four frontier models from 31B to 120B (Gemma-4-31B, Llama-4-Scout, Qwen3-32B, gpt-oss-120b), each measured with the same harness per benchmark.
+
+**Byte-per-byte on external public sets** (FLORES, BOUQuET, UD), lower is better.
+
+| Model | Params | FLORES | BOUQuET | UD |
+|---|---|---|---|---|
+| **TATlit** | **478M** | **.740** | **.777** | **.502** |
+| Tweety-7B | 7B | .757 | .800 | .544 |
+| Gemma-4-31B | 31B | .772 | .939 | .691 |
+| Goldfish-tat | 125M | .778 | .801 | **.502** |
+| mGPT-1.3B | 1.3B | .950 | 1.186 | .880 |
+| gpt-oss-120b | 120B | 1.092 | 1.164 | 1.099 |
+| Qwen3-32B | 32B | 1.144 | 1.423 | 1.159 |
+| Llama-4-Scout | 109B | 1.222 | 1.498 | 1.080 |
+
+**Byte-per-byte on in-house held-out sets** (literary, periodical), lower is better.
+
+| Model | Params | Held-out lit | Held-out per |
+|---|---|---|---|
+| **TATlit** | **478M** | **.732** | **.701** |
+| Gemma-4-31B | 31B | .790 | .705 |
+| Tweety-7B | 7B | .810 | .778 |
+| Goldfish-tat | 125M | .929 | .874 |
+| Llama-4-Scout | 109B | 1.084 | .995 |
+| Qwen3-32B | 32B | 1.105 | .979 |
+| mGPT-1.3B | 1.3B | 1.199 | 1.094 |
+| gpt-oss-120b | 120B | 2.181 | 1.857 |
+
+**TatBLiMP, morphological acceptability**, higher is better.
+
+| Model | Params | acc | acc_norm |
+|---|---|---|---|
+| **TATlit** | **478M** | **.975** | **.958** |
+| Goldfish-tat | 125M | .974 | **.958** |
+| Tweety-7B | 7B | .956 | .915 |
+| Gemma-4-31B | 31B | .924 | .839 |
+| Llama-4-Scout | 109B | .889 | .806 |
+| Qwen3-32B | 32B | .811 | .708 |
+| gpt-oss-120b | 120B | .803 | .677 |
+| mGPT-1.3B | 1.3B | .736 | .639 |
+
+On the school-knowledge benchmark TUMLU-mini the model sits near chance (0.297 against a 0.25 random baseline), which is expected for a base model with no encyclopedic training.
 
 ## Limitations
 Base model, no instruction following, no encyclopedic knowledge. Softest axis is the person/possessive system inside izafet.