Update README.md
Browse files
README.md
CHANGED
|
@@ -28,6 +28,99 @@ Crimson Dawn was trained with the Mistral Instruct template, therefore it should
|
|
| 28 |
"[INST] Prompt goes here [/INST]"
|
| 29 |
```
|
| 30 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
## Training
|
| 32 |
Training was done twice over 2 epochs each on two 2x [NVIDIA A6000 GPUs](https://www.nvidia.com/en-us/design-visualization/rtx-a6000/) using LoRA. A two-phased approach was used in which the base model was trained 2 epochs on RP data, the LoRA was then applied to base. Finally, the new modified base was trained 2 epochs on instruct, and the new instruct LoRA was applied to the modified base, resulting in what you see here.
|
| 33 |
|
|
|
|
| 28 |
"[INST] Prompt goes here [/INST]"
|
| 29 |
```
|
| 30 |
|
| 31 |
+
### Current Top Sampler Settings
|
| 32 |
+
```json
|
| 33 |
+
{
|
| 34 |
+
"temp": 1.25,
|
| 35 |
+
"temperature_last": true,
|
| 36 |
+
"top_p": 1,
|
| 37 |
+
"top_k": -1,
|
| 38 |
+
"top_a": 0,
|
| 39 |
+
"tfs": 1,
|
| 40 |
+
"epsilon_cutoff": 0,
|
| 41 |
+
"eta_cutoff": 0,
|
| 42 |
+
"typical_p": 1,
|
| 43 |
+
"min_p": 0.3,
|
| 44 |
+
"rep_pen": 1,
|
| 45 |
+
"rep_pen_range": 0,
|
| 46 |
+
"rep_pen_decay": 0,
|
| 47 |
+
"rep_pen_slope": 1,
|
| 48 |
+
"no_repeat_ngram_size": 0,
|
| 49 |
+
"penalty_alpha": 0,
|
| 50 |
+
"num_beams": 1,
|
| 51 |
+
"length_penalty": 1,
|
| 52 |
+
"min_length": 0,
|
| 53 |
+
"encoder_rep_pen": 1,
|
| 54 |
+
"freq_pen": 0,
|
| 55 |
+
"presence_pen": 0,
|
| 56 |
+
"skew": 0,
|
| 57 |
+
"do_sample": true,
|
| 58 |
+
"early_stopping": false,
|
| 59 |
+
"dynatemp": false,
|
| 60 |
+
"min_temp": 0,
|
| 61 |
+
"max_temp": 2,
|
| 62 |
+
"dynatemp_exponent": 1,
|
| 63 |
+
"smoothing_factor": 0,
|
| 64 |
+
"smoothing_curve": 1,
|
| 65 |
+
"dry_allowed_length": 2,
|
| 66 |
+
"dry_multiplier": 0,
|
| 67 |
+
"dry_base": 1.75,
|
| 68 |
+
"dry_sequence_breakers": "[\"\\n\", \":\", \"\\\"\", \"*\"]",
|
| 69 |
+
"dry_penalty_last_n": 0,
|
| 70 |
+
"add_bos_token": true,
|
| 71 |
+
"ban_eos_token": false,
|
| 72 |
+
"skip_special_tokens": true,
|
| 73 |
+
"mirostat_mode": 0,
|
| 74 |
+
"mirostat_tau": 5,
|
| 75 |
+
"mirostat_eta": 0.1,
|
| 76 |
+
"guidance_scale": 1,
|
| 77 |
+
"negative_prompt": "",
|
| 78 |
+
"grammar_string": "",
|
| 79 |
+
"json_schema": {},
|
| 80 |
+
"banned_tokens": "",
|
| 81 |
+
"sampler_priority": [
|
| 82 |
+
"temperature",
|
| 83 |
+
"dynamic_temperature",
|
| 84 |
+
"quadratic_sampling",
|
| 85 |
+
"top_k",
|
| 86 |
+
"top_p",
|
| 87 |
+
"typical_p",
|
| 88 |
+
"epsilon_cutoff",
|
| 89 |
+
"eta_cutoff",
|
| 90 |
+
"tfs",
|
| 91 |
+
"top_a",
|
| 92 |
+
"min_p",
|
| 93 |
+
"mirostat"
|
| 94 |
+
],
|
| 95 |
+
"samplers": [
|
| 96 |
+
"top_k",
|
| 97 |
+
"tfs_z",
|
| 98 |
+
"typical_p",
|
| 99 |
+
"top_p",
|
| 100 |
+
"min_p",
|
| 101 |
+
"temperature"
|
| 102 |
+
],
|
| 103 |
+
"ignore_eos_token": false,
|
| 104 |
+
"spaces_between_special_tokens": true,
|
| 105 |
+
"speculative_ngram": false,
|
| 106 |
+
"sampler_order": [
|
| 107 |
+
5,
|
| 108 |
+
6,
|
| 109 |
+
0,
|
| 110 |
+
1,
|
| 111 |
+
2,
|
| 112 |
+
3,
|
| 113 |
+
4
|
| 114 |
+
],
|
| 115 |
+
"logit_bias": [],
|
| 116 |
+
"ignore_eos_token_aphrodite": false,
|
| 117 |
+
"spaces_between_special_tokens_aphrodite": true,
|
| 118 |
+
"rep_pen_size": 0,
|
| 119 |
+
"genamt": 1024,
|
| 120 |
+
"max_length": 16384
|
| 121 |
+
}
|
| 122 |
+
```
|
| 123 |
+
|
| 124 |
## Training
|
| 125 |
Training was done twice over 2 epochs each on two 2x [NVIDIA A6000 GPUs](https://www.nvidia.com/en-us/design-visualization/rtx-a6000/) using LoRA. A two-phased approach was used in which the base model was trained 2 epochs on RP data, the LoRA was then applied to base. Finally, the new modified base was trained 2 epochs on instruct, and the new instruct LoRA was applied to the modified base, resulting in what you see here.
|
| 126 |
|