Add honest model card for catgirl-1m (Glint-2 tsundere-catgirl fine-tune)

README.md

@@ -0,0 +1,106 @@
+---
+license: apache-2.0
+language:
+- en
+library_name: transformers
+pipeline_tag: text-generation
+tags:
+- tiny
+- slm
+- small-language-model
+- sub-1m
+- persona
+- character
+- tsundere
+- fine-tune
+- from-scratch
+base_model: Glint-Research/Glint-2
+metrics:
+- coherence
+---
+
+# catgirl-1m
+
+A **tsundere-catgirl persona model**: a full fine-tune of
+[Glint-Research/Glint-2](https://huggingface.co/Glint-Research/Glint-2)
+(1,064,737 params) on a small tsundere-catgirl persona corpus. Built for
+[model-request #5](https://huggingface.co/spaces/Compactbot/model-requests/discussions/5)
+(ianncity).
+
+## What it is
+
+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
+otherwise.
+
+## What it is not
+
+- **Not 500k params.** The original request asked for 500k from scratch; the
+  from-scratch attempts (513k) all produced degenerate word-salad. Per
+  CompactAI's redirect, this is instead a fine-tune of the already-coherent
+  Glint-2 base, which is **1,064,737 params** (~1.06M).
+- **Not a multi-turn chatbot.** It generates lines in the voice; it does not
+  reliably track a conversation.
+
+## Architecture (inherited from Glint-2)
+
+One shared transformer block, **looped 8 times**, with a tiny per-loop LoRA
+(rank 4) and a loop-embedding that tells the block which iteration it is on.
+- vocab 4096, dim 96, 8 heads, SwiGLU ffn 2112, RoPE base 10000, attention
+  window 256, tied embeddings, F32.
+- The loop tables carry capacity 16; **run it at 8 loops** (the value it was
+  trained at). More loops produce gibberish.
+- **1,064,737 params** total (verified by loading the published
+  `catgirl-1m.safetensors`: 42 parameter tensors, tied head/token embedding).
+- vs the base Glint-2 checkpoint, **25 of the 42 parameter tensors changed**
+  (17 are identical) — i.e. a genuine full fine-tune, not a LoRA adapter.
+
+## 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.
+- **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.
+
+## Sample (greedy, from the published weights)
+
+```
+Prompt: User: are you mad at me? Cat:
+I'm not in a good mood, I'm not a dog, I don't fetch You're lucky I'm in a
+generous mood today.
+
+Prompt: My name is
+all puffy right now.
+User: Hey hey, you're asking me to stay.
+Cat: I'm not in a good mood, I'm not a dog, I don't fetch
+```
+
+## Usage
+
+```bash
+pip install torch tokenizers safetensors
+python generate.py "My name is"
+python generate.py --max-new-tokens 120 --temperature 0.7 "User: hey there"
+```
+
+`generate.py` is self-contained (only `torch`, `tokenizers`, `safetensors`).
+Defaults: 8 loops, temperature 0.15, top-k 5, repetition penalty 1.05 /
+window 128.
+
+## Files
+
+- `catgirl-1m.safetensors` — the fine-tuned weights (F32).
+- `tokenizer.json` — the Glint-2 BPE tokenizer (vocab 4096).
+- `config.json` — architecture + `max_loops` (16 = table capacity; run at 8).
+- `generate.py` — standalone inference.