Add honest model card (corrected training numbers + known limitation)

README.md

@@ -6,6 +6,7 @@ library_name: transformers
 pipeline_tag: text-generation
 tags:
 - tiny
+- tiny-lm
 - slm
 - small-language-model
 - sub-1m
@@ -13,7 +14,6 @@ tags:
 - character
 - tsundere
 - fine-tune
-- from-scratch
 base_model: Glint-Research/Glint-2
 metrics:
 - coherence
@@ -33,7 +33,7 @@ A tiny looped transformer that generates **in-voice tsundere-catgirl lines**:
 the grudging-then-soft cadence, the "it's not like I wanted to!" phrasing,
 catgirl mannerisms. It is a **style model, not a conversationalist** — it
 produces short in-voice lines and does not hold coherent multi-turn dialogue
-or track long context. That is well beyond 1M params and I will not pretend
+or track long context. That is well beyond ~1M params and I will not pretend
 otherwise.
 
 ## What it is not
@@ -61,17 +61,30 @@ One shared transformer block, **looped 8 times**, with a tiny per-loop LoRA
 ## Training
 
 - **Data:** a 216,144-byte / 3,933-line tsundere-catgirl persona corpus
-  (User:/Cat: dialogue in the target voice).
-- **Run:** 5000 steps, batch 16, seq 128, lr 1e-4, cosine decay with 10%
-  warmup.
+  (User:/Cat: dialogue in the target voice) = **82,695 tokens** with the
+  Glint-2 BPE tokenizer, packed into 323 sequences of 256.
+- **Run:** 2000 steps, batch 8, seq 256, lr 2e-4, cosine decay with 10% warmup
+  (≈ 662k tokens seen).
 - **Device:** NVIDIA RTX 5090 (32 GB).
-- **Ship gate:** 10 fresh greedy samples — **0/10 degenerate**, all coherent
-  and in-voice.
-
-> **No validation-loss number is reported on purpose.** The trainer's
-> "held-out" slice was drawn from the same training sequence pool, so its val
-> loss is not a real held-out number and is not quoted here. Quality is
-> judged by the samples, not by a val loss.
+- **Ship gate:** 10 fresh greedy samples — **0/10 degenerate** by the
+  6-word-ngram detector, all in-voice.
+
+> **No validation-loss number is reported on purpose.** The trainer produced
+> no finite held-out val loss for this run (`best_val` stayed `inf`), so there
+> is no real held-out number to quote. Quality is judged by the samples, not by
+> a val loss.
+
+## Known limitation (honest)
+
+The corpus is built around a small set of "User:" turn templates — most
+prominently *"you're asking me to [verb]"*, which appears ~696 times across
+only ~8 distinct verbs (purr, stay, nap, sit, meow, help, play, come). The
+model faithfully learned this: in free generation it tends to **echo that
+template in the User turns** (with different verbs), even when the prompt did
+not contain it. The **Cat replies are varied and on-persona**; the template
+locking is in the User side. As a "give it a User line, get a tsundere Cat
+reply" persona model it works well; as an open-ended generator it leans on the
+template. A larger, more diverse corpus would fix this.
 
 ## Sample (greedy, from the published weights)