Commit 8d598a5
Fix Qwen fallback recommendation in dash-ocr patterns memo
Per dash-ocr-pipeline's current production code (`structured_stage12.py`,
`retry_silent_empty_checkpoint.py`), the silent-empty fallback uses the
*same* Gemini 2.0 Flash model with a permissive V1 prompt, not a switch
to Qwen 2.5-72B. The two-stage Gemini→Qwen pairer path was retired:
single-call Gemini Flash F1 = 0.965 beat the Qwen pairer F1 = 0.950
in production validation 2026-05-06.
Updated:
- §3.2 budget: T2 V1 fallback uses Gemini Flash, not Qwen (saved ~$0.01)
- §3.2 budget: T3 LLM verification uses Gemini Flash, not Qwen
- §4 defaults: T2 fallback row corrected
- §7 Q2 implication: simplified to single-model V2/V1 prompt pair
Also notes that PR #3 has already been updated with the corrected Q2
Decision (commit b299269) so the two docs now agree.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>1 parent ba487af commit 8d598a5
1 file changed
Lines changed: 9 additions & 11 deletions
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
183 | 183 | | |
184 | 184 | | |
185 | 185 | | |
186 | | - | |
187 | | - | |
188 | | - | |
189 | | - | |
| 186 | + | |
| 187 | + | |
| 188 | + | |
| 189 | + | |
190 | 190 | | |
191 | 191 | | |
192 | | - | |
| 192 | + | |
193 | 193 | | |
194 | | - | |
| 194 | + | |
195 | 195 | | |
196 | 196 | | |
197 | 197 | | |
| |||
203 | 203 | | |
204 | 204 | | |
205 | 205 | | |
206 | | - | |
| 206 | + | |
207 | 207 | | |
208 | 208 | | |
209 | 209 | | |
| |||
241 | 241 | | |
242 | 242 | | |
243 | 243 | | |
244 | | - | |
| 244 | + | |
245 | 245 | | |
246 | 246 | | |
247 | | - | |
248 | | - | |
249 | | - | |
| 247 | + | |
250 | 248 | | |
251 | 249 | | |
252 | 250 | | |
| |||
0 commit comments