Test think prefill with pinned SFT epoch 3 revision

Dockerfile

@@ -16,6 +16,7 @@ ENV HOME=/home/user \
     PATH=/opt/venv/bin:$PATH \
     MODEL_NAME=tr-hash-moe-200m \
     MODEL_REPO=AETHORIA-AI/TR-HASH-MoE-200M-160B-SFT \
+    MODEL_REVISION=e5a599feda17ce9ffa8a0d094fd591b06a4d16f5 \
     CONTEXT_COMPACT_TOKENS=1024 \
     PORT=7860
 

README.md

@@ -6,27 +6,31 @@ colorTo: green
 sdk: docker
 app_port: 7860
 pinned: false
-license: cc-by-nc-4.0
+license: apache-2.0
 models:
   - AETHORIA-AI/TR-HASH-MoE-200M-160B-SFT
 ---
 
 # TR-Hash Tiny · TR-HASH MoE 200M Full SFT
 
-OpenAI-compatible chat API serving the PIQA-selected epoch-2 root weights from
+OpenAI-compatible chat API serving the PIQA-selected epoch-3 root weights from
 [`AETHORIA-AI/TR-HASH-MoE-200M-160B-SFT`](https://huggingface.co/AETHORIA-AI/TR-HASH-MoE-200M-160B-SFT)
 with [`Complexity-ML/TR-Hash-i64`](https://github.com/Complexity-ML/TR-Hash-i64).
 
 This is a full-parameter SFT, not a LoRA adapter. All 201.2M parameters were
-trained for three epochs on the audited Luciole 16-way mixture. Epoch 2 scores
-**68.82% PIQA accuracy** and **69.31% length-normalized accuracy** on the full
-1,838-example validation set.
+trained for three epochs on the audited 300K SFT v2 mixture. Epoch 3 scores
+**68.01% PIQA accuracy** and **69.21% length-normalized accuracy** on the full
+1,838-example validation set in the MLX FP16 cross-check.
 
 The launcher intentionally downloads only the release files at the model-repo
 root. The three large resumable `step_*` training folders are excluded. The
 bundled `chat_template.jinja` and tokenizer are loaded beside the single root
 `model.safetensors`.
 
+This experimental deployment pre-fills `<think>` in the assistant generation
+prompt. The model is expected to close it with `</think>` and then emit a
+`<final>` answer, but this behavior is under live evaluation.
+
 ```bash
 curl https://pacific-i64-tr-hash-tiny.hf.space/v1/chat/completions \
   -H "Content-Type: application/json" \
@@ -34,9 +38,9 @@ curl https://pacific-i64-tr-hash-tiny.hf.space/v1/chat/completions \
     "model": "tr-hash-moe-200m",
     "messages": [{"role": "user", "content": "Explain why 17 × 23 − 14 equals 377."}],
     "max_tokens": 1024,
-    "temperature": 0.4,
-    "top_k": 30,
-    "top_p": 0.85,
+    "temperature": 0.3,
+    "top_k": 20,
+    "top_p": 0.95,
     "repetition_penalty": 1.1,
     "stream": false
   }'

app.py

@@ -9,7 +9,6 @@ from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
 from pathlib import Path
 
 from huggingface_hub import snapshot_download
-from huggingface_hub.errors import LocalEntryNotFoundError
 
 
 def _available_cpus() -> int:
@@ -39,6 +38,7 @@ model_name = os.environ["MODEL_NAME"]
 model_repo = os.environ.get(
     "MODEL_REPO", "AETHORIA-AI/TR-HASH-MoE-200M-160B-SFT"
 )
+model_revision = os.environ.get("MODEL_REVISION") or None
 model_dir_override = os.environ.get("MODEL_DIR")
 port = int(os.environ.get("PORT", "7860"))
 MODEL_FILES = [
@@ -82,21 +82,15 @@ def _resolve_checkpoint() -> str:
     thread = threading.Thread(target=server.serve_forever, daemon=True)
     thread.start()
     print(
-        f"Resolving {model_repo} from the preloaded Hugging Face cache.",
+        f"Resolving {model_repo} at revision {model_revision or 'main'}.",
         flush=True,
     )
     try:
-        try:
-            checkpoint_dir = snapshot_download(
-                repo_id=model_repo,
-                allow_patterns=MODEL_FILES,
-                local_files_only=True,
-            )
-        except LocalEntryNotFoundError:
-            checkpoint_dir = snapshot_download(
-                repo_id=model_repo,
-                allow_patterns=MODEL_FILES,
-            )
+        checkpoint_dir = snapshot_download(
+            repo_id=model_repo,
+            revision=model_revision,
+            allow_patterns=MODEL_FILES,
+        )
     finally:
         server.shutdown()
         server.server_close()