Add TestGeniy v6 dialogue SFT candidate metadata
Browse files- dialogue_sft_v6/README.md +20 -0
dialogue_sft_v6/README.md
ADDED
|
@@ -0,0 +1,20 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# TestGeniy v6 Dialogue SFT
|
| 2 |
+
|
| 3 |
+
This folder contains the accepted dialogue-focused SFT candidate for TestGeniy. It is uploaded separately and does not replace the repository root model files.
|
| 4 |
+
|
| 5 |
+
## Provenance
|
| 6 |
+
|
| 7 |
+
- Base: `Asilarkness/testgeniy@116679b75b635dcbb0d21d068b74fcafc599947b`
|
| 8 |
+
- Real data only; no synthetic answers and no MTP
|
| 9 |
+
- `OpenAssistant/oasst1`: human-authored and human-ranked dialogue
|
| 10 |
+
- `nvidia/HelpSteer2`: human quality annotations and human preference data
|
| 11 |
+
|
| 12 |
+
## Training
|
| 13 |
+
|
| 14 |
+
Full-parameter SFT, AdamW, 420 steps, micro-batch 2, gradient accumulation 8, learning rate `5e-7`, maximum sequence length 768.
|
| 15 |
+
|
| 16 |
+
## Evaluation
|
| 17 |
+
|
| 18 |
+
On the fixed 100-example suite, the candidate changed the aggregate from **19.75** to **21.5**: GSM8K 20>24, MATH-500 4>7, ARC-Challenge 21>26, FOLIO 34>29. It is therefore accepted as a dialogue-focused candidate, but not as an unconditional universal replacement: FOLIO and separate human dialogue preference holdouts regressed.
|
| 19 |
+
|
| 20 |
+
See `manifest.json` for exact metadata.
|