fix(v5): proper space tokenization - replace UNK injection with explicit SPACE_ID

fix: hidden_size 320->384 to be divisible by 12 heads (head_dim=32); ff_dim 768->1024

feat: v4 upgrade - 35M params, vocab=800, 560+ QA pairs, 10 domains

feat: v3 upgrade - vocab 2000, seq_len 256, answer-only loss training

feat: QA separator tokens - <|q|>/<|a|> special tokens for better generation

feat: v2 upgrade - 12M params, BPE tokenizer, 200+ QA pairs, 100 epochs max

fix: causal_mask slice preserve (1,1,T,T) shape for head broadcasting

fix: remove attention_mask arg - FusionMini in app.py uses internal causal mask

fix: pass causal attention_mask in chat_fn to fix tensor size mismatch

debug: wrap train/chat in try/except to show error in UI

fix: replace LinePlot with plain Textbox for gradio 4.44 compat

fix: remove server_name/port from launch(), use LinePlot for gradio 4.44 compat

fix: lock jinja2/starlette for gradio 4.44.0 compat + launch fallback for localhost bug

fix: pin gradio==4.40.0 + huggingface_hub==0.25.2 to avoid 4.44.0 jinja2/launch bugs

fix: remove explicit server_name/port in demo.launch() for HF Spaces compatibility

fix: add HfFolder compat shim for huggingface_hub >=1.0 + gradio 4.44

fix: remove gradio pin, use HF Spaces pre-installed version

fix: pin gradio==4.40.0 to resolve HF Spaces dependency conflict

fix: pin gradio<4.44.0 to avoid HfFolder removal in huggingface_hub

fix: unpin gradio version for HF Spaces compatibility

merge: resolve conflicts, keep local requirements.txt with gradio==4.12.0, remove spaces/ subdir

fix: pin gradio==4.12.0 to resolve dependency conflict

fix: pin gradio<4.40.0 to avoid HfFolder import error

fix: correct tqdm version from 65.0.0 to 4.65.0

fix: move Space files to repo root for HF config detection

feat: add HF Spaces demo with Gradio UI (SBLA + Thinking Dial)

feat: Hugging Face Spaces — Gradio demo with training + chat

fix: F5-1 — initialize self._tokenizer before _ensure_tokenizer()

fix: F4-1/2/3 — three crash-on-run bugs

fix(audit): C1/H1/H3/H4 审计报告修复

fix: D19 add upper bound constraints to requirements.txt

fix: D16 migrate 4 test files from print-based to pytest convention

fix: audit v2 P1 fixes - D8 Ollama SBLA warning, D13 metrics deprecation

fix: audit v2 P0/P1 fixes (D1/D2/D3/D7/D9/D11/D14)

fix: distillation script uses FusionModel + add GRPO training script

fix: comprehensive audit fixes (ONNX/export, benchmark, deprecated APIs, attention unification)

fix: replace random embeddings in MoverScore with deterministic hash

fix: move FusionMini @classmethods from FusionMiniConfig to FusionMini class

fix: test suite + bilingual filter + TensorRT/OpenVINO export rewrite

refactor: unify Thinking Dial to architecture-level only, deprecate text injection

fix: N9 position_ids signatures + M2 THINK_END token split

chore: remove temp verification scripts

fix: comprehensive audit fixes - Thinking Dial unification, deployment scripts, README, bilingual filter, data download

fix: from_pretrained weight loading (HF 5.x compat) + _init_weights + tied_weights_keys

fix: CausalLMOutputWithPast return + audit verification

feat: setup.py, quick-start guide, architecture doc, model registry

fix: short-sequence benchmarks + ThinkingDial vs vanilla experiments

feat: GSM8K reward function + GRPO evaluation pipeline (v2.1.1)

v2.2.0: end-to-end pipeline validation + GRPO no-tokenizer fallback

v2.1.0: fix N25 off-by-one, S2 reward_fn string safety