Update README.md
Browse files
README.md
CHANGED
|
@@ -24,13 +24,59 @@ license: apache-2.0
|
|
| 24 |
|
| 25 |
|
| 26 |
|
|
|
|
| 27 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
|
| 29 |
### Training details
|
| 30 |
|
| 31 |
<details>
|
| 32 |
<summary>Spoiler warning</summary>
|
| 33 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
The following YAML configuration was used to produce this model:
|
| 35 |
|
| 36 |
```yaml
|
|
|
|
| 24 |
|
| 25 |
|
| 26 |
|
| 27 |
+
A generalist model, with some reasoning capabilities and some multi-lang support.
|
| 28 |
|
| 29 |
+
Supported languages:
|
| 30 |
+
- French
|
| 31 |
+
- German
|
| 32 |
+
- Spanish
|
| 33 |
+
- Italian
|
| 34 |
+
- Portuguese
|
| 35 |
+
- Russian
|
| 36 |
+
- Chinese
|
| 37 |
+
- Japanese
|
| 38 |
+
|
| 39 |
+
This model is trained in FFT based on unreleased model merge (uses same models as [Retreatcost/KansenSakura-Erosion-RP-12b](https://huggingface.co/Retreatcost/KansenSakura-Erosion-RP-12b), credits to all original model authors.), using in-progress dateset, that I am creating for another project.
|
| 40 |
+
|
| 41 |
+
Training stats can be found in "Training metrics" tab.
|
| 42 |
+
|
| 43 |
+
Reasoning should work out of the box most of the time with occasional replies without it.
|
| 44 |
+
For absolute consistency you can prefill model responses with "< think >\n" (think tag without spaces, line break is preferred).
|
| 45 |
+
|
| 46 |
+
## Intended use
|
| 47 |
+
|
| 48 |
+
- General conversations, chatting.
|
| 49 |
+
- Co-writing, brainstorming.
|
| 50 |
+
- Short roleplaying
|
| 51 |
+
|
| 52 |
+
## Inference Tips
|
| 53 |
+
|
| 54 |
+
1. **Temperature**: 0.7 (0.6 - 0.8 range should work fine)
|
| 55 |
+
2. **Repetition Penalty**: 1.05
|
| 56 |
+
3. **TOP_P**: 0.90
|
| 57 |
+
4. **TOP_K**: 0 (disable)
|
| 58 |
+
5. **MIN_P**: 0.025
|
| 59 |
+
6. **Template Format**: ChatML
|
| 60 |
+
7. **Max Output**: 1024 (Due to additional reasoning budget I suggest giving the model at least 768 tokens, preferrably over 1K, but usually it rarely outputs answers longer than 1.35K).
|
| 61 |
+
6. **Context Management**: 8K
|
| 62 |
+
|
| 63 |
+
I haven't really tested or trained the model for long context, so it will probably break earlier than regular models.
|
| 64 |
+
You can set a higher context, for example 16K, 24K or 32K, but I don't guarantee how it will behave.
|
| 65 |
|
| 66 |
### Training details
|
| 67 |
|
| 68 |
<details>
|
| 69 |
<summary>Spoiler warning</summary>
|
| 70 |
|
| 71 |
+
I trained 2 variants of the model:
|
| 72 |
+
- with unrolled turns (each turn in separate sample)
|
| 73 |
+
- with regular turns (all turns in single sample)
|
| 74 |
+
|
| 75 |
+
Unrolled turns teach local attention much better and train faster, but generalize worse for multi-turn (LA, Local attention).
|
| 76 |
+
Regular turns have much better multi-turn generalisation, but they tend to memorize instead of training new capabilities. (GA, Global attention)
|
| 77 |
+
|
| 78 |
+
Then I merged these training runs in a pattern 4:1, similar to how Gemma models have layered SWA and GA.
|
| 79 |
+
|
| 80 |
The following YAML configuration was used to produce this model:
|
| 81 |
|
| 82 |
```yaml
|