HatemSaadallah commited on
Commit
00fd2c2
·
verified ·
1 Parent(s): cac55c3

Card: the final system, with its disclosure and its own comparison with the base model

Browse files
Files changed (1) hide show
  1. README.md +7 -3
README.md CHANGED
@@ -226,10 +226,11 @@ a stronger or human-corrected teacher for evidence judgement, is the open direct
226
  - **Exact reproduction from the public data.** Since 11 November the public dossier is byte-identical to the one the
227
  evaluated runs read (the release of 8 November had cut excerpts to one short quote).
228
 
229
- ## The final system: the version 3 adapter (`exploratory/v3_adapter.gguf`)
230
 
231
  A second adapter, continued from this one on rows from a second teacher (DeepSeek V4-Pro; 3,611 rows from 144
232
- replay-exact investigations). On its own it did not pass its development adoption rule; it is the reasoner of the final system. It is
 
233
  published so the two-adapter system described in the write-up ("dig then reason": this card's adapter gathers the
234
  evidence, version 3 reasons) can be run by others. File: GGUF for llama.cpp, sha256 `ea919380…e1443`, the conversion
235
  the final system's runs used. Blind scores, mean of two judges (Claude Code and Codex `gpt-6-sol`), base / version 1 /
@@ -238,7 +239,10 @@ version 3: test set 2 0.56 / 1.06 / 1.48; final test 0.62 / 1.46 / 1.33. Dig the
238
  ## The final system: decoding-time guards (the app's default)
239
 
240
  They change what the model can produce at each step, not its weights; the version 1 numbers above use none of
241
- them. With the two-adapter system ("dig then reason"):
 
 
 
242
  - **Loop guards**: a source already weighed for a theory and a search already run are not offered again, a quote is
243
  picked from the source's own sentences, an exact repeat is drawn again before it counts, and a step-budget guard.
244
  Finished runs: development 28 → 45 of 45, test set 2 43 → 69 of 72, final test 27 → 45 of 48. Blind quality, mean of
 
226
  - **Exact reproduction from the public data.** Since 11 November the public dossier is byte-identical to the one the
227
  evaluated runs read (the release of 8 November had cut excerpts to one short quote).
228
 
229
+ ## The final system: the version 3 adapter (`v3_adapter.gguf`)
230
 
231
  A second adapter, continued from this one on rows from a second teacher (DeepSeek V4-Pro; 3,611 rows from 144
232
+ replay-exact investigations). On its own it did not pass its development adoption rule; as the reasoner of dig then reason it did (+0.17 [0.02,
233
+ 0.35] over version 1 on development), and that is how the final system uses it. It is
234
  published so the two-adapter system described in the write-up ("dig then reason": this card's adapter gathers the
235
  evidence, version 3 reasons) can be run by others. File: GGUF for llama.cpp, sha256 `ea919380…e1443`, the conversion
236
  the final system's runs used. Blind scores, mean of two judges (Claude Code and Codex `gpt-6-sol`), base / version 1 /
 
239
  ## The final system: decoding-time guards (the app's default)
240
 
241
  They change what the model can produce at each step, not its weights; the version 1 numbers above use none of
242
+ them. They were built after both frozen tests, partly from the failures in their runs (the loop guards from the
243
+ refused calls of the final-test runs), and measured on the same cases, so version 1's numbers stay the run-once
244
+ result. Against the base model, the final system scores +1.02 [0.92, 1.10] on the first test's cases and +1.11
245
+ [1.01, 1.20] on the second's, judged by the same two judges in separate batches. With the two-adapter system ("dig then reason"):
246
  - **Loop guards**: a source already weighed for a theory and a search already run are not offered again, a quote is
247
  picked from the source's own sentences, an exact repeat is drawn again before it counts, and a step-budget guard.
248
  Finished runs: development 28 → 45 of 45, test set 2 43 → 69 of 72, final test 27 → 45 of 48. Blind quality, mean of