Revert to Qwen2.5-72B-Instruct; 8B failed prompt reasoning test

Llama-3.1-8B printed literal [name] placeholder and conflated
first-time vs returning user greeting — not capable of following
the 9K token v2 counselling prompt reliably. Qwen2.5-72B-Instruct
is the confirmed working model on Novita.

Generated with Devin (https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>

CHANGELOG.md

@@ -24,11 +24,10 @@ Dates reflect the git commit date. Commit hashes are abbreviated to 7 characters
 - **LLM switched to `meta-llama/Llama-3.1-8B-Instruct` on Novita** (`bee8a73`)
   - `Llama-3.3-70B-Instruct` was dropped by Novita (confirmed: "Model not supported by provider")
   - `Llama-3.1-8B-Instruct` is confirmed available on Novita — currently under user testing
-  - **PENDING ASSESSMENT:** 8B is significantly smaller than the previous 70B model.
-    If response quality is insufficient for Sanyu's nuanced counselling use case,
-    the planned fallback is `Qwen/Qwen2.5-72B-Instruct` (also confirmed on Novita,
-    previously tested and working — see commit `9a39c3f`).
-  - Decision point: after user testing, either keep 8B or revert to Qwen2.5-72B
+  - **ASSESSED & REJECTED:** 8B model printed literal `[name]` placeholder and
+    conflated first-time vs returning user greeting scripts — insufficient reasoning
+    capacity for the ~9K token v2 prompt with complex conditional logic.
+  - **DECISION: reverted to `Qwen/Qwen2.5-72B-Instruct`** as the permanent model.
 
 ---
 

app.py

@@ -941,7 +941,7 @@ def respond(message, history):
     for attempt in range(3):
         try:
             response = client.chat_completion(
-                model="meta-llama/Llama-3.1-8B-Instruct",  # confirmed available on Novita; fallback is Qwen/Qwen2.5-72B-Instruct
+                model="Qwen/Qwen2.5-72B-Instruct",  # 8B too small for complex prompt; Qwen2.5-72B confirmed working on Novita
                 messages=messages,
                 max_tokens=512,
             )