Text Generation
MLX
Safetensors
English
small-language-model
from-scratch
conversational
apple-silicon
research
Instructions to use OpenSML/Solstice-1-175M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use OpenSML/Solstice-1-175M with MLX:
# Make sure mlx-lm is installed # pip install --upgrade mlx-lm # Generate text with mlx-lm from mlx_lm import load, generate model, tokenizer = load("OpenSML/Solstice-1-175M") prompt = "Write a story about Einstein" messages = [{"role": "user", "content": prompt}] prompt = tokenizer.apply_chat_template( messages, add_generation_prompt=True ) text = generate(model, tokenizer, prompt=prompt, verbose=True) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- MLX LM
How to use OpenSML/Solstice-1-175M with MLX LM:
Generate or start a chat session
# Install MLX LM uv tool install mlx-lm # Interactive chat REPL mlx_lm.chat --model "OpenSML/Solstice-1-175M"
Run an OpenAI-compatible server
# Install MLX LM uv tool install mlx-lm # Start the server mlx_lm.server --model "OpenSML/Solstice-1-175M" # Calling the OpenAI-compatible server with curl curl -X POST "http://localhost:8080/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "OpenSML/Solstice-1-175M", "messages": [ {"role": "user", "content": "Hello"} ] }' - Atomic Chat
Rename model to Solstice-1-175M
Browse files- CITATION.cff +3 -3
- FILE_MANIFEST.json +52 -52
- README.md +13 -11
- REPOSITORIES.json +3 -3
- RUN_GUIDE.md +3 -3
- SHA256SUMS +27 -27
- SOURCE_NOTICE.md +1 -1
- THIRD_PARTY_NOTICES.md +2 -2
- TRAINING_AND_REPRODUCIBILITY.md +1 -1
- config.json +2 -2
- docs/DATA.md +1 -1
- docs/EVALUATION.md +5 -5
- docs/TECHNICAL_REPORT.md +3 -3
- docs/figures/benchmark-comparison.png +2 -2
- docs/figures/benchmark-comparison.svg +220 -144
- docs/figures/chat-diagnostics.png +2 -2
- docs/figures/chat-diagnostics.svg +103 -47
- docs/figures/cooldown-detail.png +2 -2
- docs/figures/cooldown-detail.svg +112 -121
- docs/figures/ifeval-comparison.png +2 -2
- docs/figures/ifeval-comparison.svg +117 -56
- docs/figures/pretraining-overview.png +2 -2
- docs/figures/pretraining-overview.svg +100 -109
- inference_native/__init__.py +1 -1
- provenance/LINEAGE.json +1 -1
- scripts/plot_results.py +7 -7
- scripts/plot_training.py +1 -1
- scripts/verify_package.py +1 -1
CITATION.cff
CHANGED
|
@@ -1,10 +1,10 @@
|
|
| 1 |
cff-version: 1.2.0
|
| 2 |
-
message: "If you use Solstice-1, cite this model and its recorded release revision."
|
| 3 |
-
title: "Solstice-1: a 175M English chat research model"
|
| 4 |
authors:
|
| 5 |
- family-names: Zebrowski
|
| 6 |
given-names: William
|
| 7 |
type: software
|
| 8 |
version: "1.0.0"
|
| 9 |
license: Apache-2.0
|
| 10 |
-
url: "https://huggingface.co/OpenSML/Solstice-1"
|
|
|
|
| 1 |
cff-version: 1.2.0
|
| 2 |
+
message: "If you use Solstice-1-175M, cite this model and its recorded release revision."
|
| 3 |
+
title: "Solstice-1-175M: a 175M English chat research model"
|
| 4 |
authors:
|
| 5 |
- family-names: Zebrowski
|
| 6 |
given-names: William
|
| 7 |
type: software
|
| 8 |
version: "1.0.0"
|
| 9 |
license: Apache-2.0
|
| 10 |
+
url: "https://huggingface.co/OpenSML/Solstice-1-175M"
|
FILE_MANIFEST.json
CHANGED
|
@@ -10,8 +10,8 @@
|
|
| 10 |
"sha256": "f09ed498cc42b482dae88d9fb5026bf74e1d8b9187f9ea301aeda0db76d6c859"
|
| 11 |
},
|
| 12 |
"CITATION.cff": {
|
| 13 |
-
"bytes":
|
| 14 |
-
"sha256": "
|
| 15 |
},
|
| 16 |
"LICENSE": {
|
| 17 |
"bytes": 11358,
|
|
@@ -22,32 +22,32 @@
|
|
| 22 |
"sha256": "0e6865f1a81cfd69ca48e6987658854ce0c0d23e91d72ae545393cd5fa032375"
|
| 23 |
},
|
| 24 |
"README.md": {
|
| 25 |
-
"bytes":
|
| 26 |
-
"sha256": "
|
| 27 |
},
|
| 28 |
"REPOSITORIES.json": {
|
| 29 |
-
"bytes":
|
| 30 |
-
"sha256": "
|
| 31 |
},
|
| 32 |
"RUN_GUIDE.md": {
|
| 33 |
-
"bytes":
|
| 34 |
-
"sha256": "
|
| 35 |
},
|
| 36 |
"SOURCE_NOTICE.md": {
|
| 37 |
-
"bytes":
|
| 38 |
-
"sha256": "
|
| 39 |
},
|
| 40 |
"THIRD_PARTY_NOTICES.md": {
|
| 41 |
-
"bytes":
|
| 42 |
-
"sha256": "
|
| 43 |
},
|
| 44 |
"TRAINING_AND_REPRODUCIBILITY.md": {
|
| 45 |
-
"bytes":
|
| 46 |
-
"sha256": "
|
| 47 |
},
|
| 48 |
"config.json": {
|
| 49 |
-
"bytes":
|
| 50 |
-
"sha256": "
|
| 51 |
},
|
| 52 |
"configs/pretrain_20261022_121457_b3d5f398/corpus.json": {
|
| 53 |
"bytes": 1356,
|
|
@@ -126,60 +126,60 @@
|
|
| 126 |
"sha256": "ae708eca3c09d5a27eab3fddbe4106d0658787beebc9393de57f75234480a284"
|
| 127 |
},
|
| 128 |
"docs/DATA.md": {
|
| 129 |
-
"bytes":
|
| 130 |
-
"sha256": "
|
| 131 |
},
|
| 132 |
"docs/EVALUATION.md": {
|
| 133 |
-
"bytes":
|
| 134 |
-
"sha256": "
|
| 135 |
},
|
| 136 |
"docs/EXAMPLES.md": {
|
| 137 |
"bytes": 1117,
|
| 138 |
"sha256": "7a818c365f1c047b349479829396ba5cd841ca8d2f04fa14fc1489083a1fd4b3"
|
| 139 |
},
|
| 140 |
"docs/TECHNICAL_REPORT.md": {
|
| 141 |
-
"bytes":
|
| 142 |
-
"sha256": "
|
| 143 |
},
|
| 144 |
"docs/figures/benchmark-comparison.png": {
|
| 145 |
-
"bytes":
|
| 146 |
-
"sha256": "
|
| 147 |
},
|
| 148 |
"docs/figures/benchmark-comparison.svg": {
|
| 149 |
-
"bytes":
|
| 150 |
-
"sha256": "
|
| 151 |
},
|
| 152 |
"docs/figures/chat-diagnostics.png": {
|
| 153 |
-
"bytes":
|
| 154 |
-
"sha256": "
|
| 155 |
},
|
| 156 |
"docs/figures/chat-diagnostics.svg": {
|
| 157 |
-
"bytes":
|
| 158 |
-
"sha256": "
|
| 159 |
},
|
| 160 |
"docs/figures/cooldown-detail.png": {
|
| 161 |
-
"bytes":
|
| 162 |
-
"sha256": "
|
| 163 |
},
|
| 164 |
"docs/figures/cooldown-detail.svg": {
|
| 165 |
-
"bytes":
|
| 166 |
-
"sha256": "
|
| 167 |
},
|
| 168 |
"docs/figures/ifeval-comparison.png": {
|
| 169 |
-
"bytes":
|
| 170 |
-
"sha256": "
|
| 171 |
},
|
| 172 |
"docs/figures/ifeval-comparison.svg": {
|
| 173 |
-
"bytes":
|
| 174 |
-
"sha256": "
|
| 175 |
},
|
| 176 |
"docs/figures/pretraining-overview.png": {
|
| 177 |
-
"bytes":
|
| 178 |
-
"sha256": "
|
| 179 |
},
|
| 180 |
"docs/figures/pretraining-overview.svg": {
|
| 181 |
-
"bytes":
|
| 182 |
-
"sha256": "
|
| 183 |
},
|
| 184 |
"evaluation/README.md": {
|
| 185 |
"bytes": 1816,
|
|
@@ -250,8 +250,8 @@
|
|
| 250 |
"sha256": "e8f5f334901ee37218c0cf0273cc064af24ff00fb6535ca4c5f2f35843e3b410"
|
| 251 |
},
|
| 252 |
"inference_native/__init__.py": {
|
| 253 |
-
"bytes":
|
| 254 |
-
"sha256": "
|
| 255 |
},
|
| 256 |
"inference_native/chat_format.py": {
|
| 257 |
"bytes": 4129,
|
|
@@ -278,8 +278,8 @@
|
|
| 278 |
"sha256": "1e3598d6bdf2b16b4b00b3504cf0fd3c3484b69ab6efe97d3f790c437e85c32f"
|
| 279 |
},
|
| 280 |
"provenance/LINEAGE.json": {
|
| 281 |
-
"bytes":
|
| 282 |
-
"sha256": "
|
| 283 |
},
|
| 284 |
"provenance/PRETRAINING_PHASES.json": {
|
| 285 |
"bytes": 41733,
|
|
@@ -514,20 +514,20 @@
|
|
| 514 |
"sha256": "cac0e352b21365bfb4fef9f5c265aa36426552bb47851be5fdb5ff585648f69a"
|
| 515 |
},
|
| 516 |
"scripts/plot_results.py": {
|
| 517 |
-
"bytes":
|
| 518 |
-
"sha256": "
|
| 519 |
},
|
| 520 |
"scripts/plot_training.py": {
|
| 521 |
-
"bytes":
|
| 522 |
-
"sha256": "
|
| 523 |
},
|
| 524 |
"scripts/verify_inference.py": {
|
| 525 |
"bytes": 4559,
|
| 526 |
"sha256": "002b4e552d3d625f96d77d90a0e3195e53e5c4358f64fe3a042792d08b232cd3"
|
| 527 |
},
|
| 528 |
"scripts/verify_package.py": {
|
| 529 |
-
"bytes":
|
| 530 |
-
"sha256": "
|
| 531 |
},
|
| 532 |
"tokenizer/corpus.json": {
|
| 533 |
"bytes": 1549,
|
|
|
|
| 10 |
"sha256": "f09ed498cc42b482dae88d9fb5026bf74e1d8b9187f9ea301aeda0db76d6c859"
|
| 11 |
},
|
| 12 |
"CITATION.cff": {
|
| 13 |
+
"bytes": 338,
|
| 14 |
+
"sha256": "097c833a247c878ebc6e4b070e861e1637cf90bd25b5206fe7f3dc2bbcd0e3a4"
|
| 15 |
},
|
| 16 |
"LICENSE": {
|
| 17 |
"bytes": 11358,
|
|
|
|
| 22 |
"sha256": "0e6865f1a81cfd69ca48e6987658854ce0c0d23e91d72ae545393cd5fa032375"
|
| 23 |
},
|
| 24 |
"README.md": {
|
| 25 |
+
"bytes": 12672,
|
| 26 |
+
"sha256": "dc72b70e5db4fdf9e7a5b8307f7801c6de12e4ba45555349bd0413438ad5333a"
|
| 27 |
},
|
| 28 |
"REPOSITORIES.json": {
|
| 29 |
+
"bytes": 738,
|
| 30 |
+
"sha256": "572c3931bd1b8278208873863148b2dd99cb9882accc431fa52143d93dac0d5d"
|
| 31 |
},
|
| 32 |
"RUN_GUIDE.md": {
|
| 33 |
+
"bytes": 3169,
|
| 34 |
+
"sha256": "09554010710c463910e392079ab69ed5de4314321afb7a260a6ff4ebe9160ce5"
|
| 35 |
},
|
| 36 |
"SOURCE_NOTICE.md": {
|
| 37 |
+
"bytes": 1064,
|
| 38 |
+
"sha256": "819ca2766a86d1ff4e8a5265a9d03d4ddc315a9256e2c008626c9e0df462489b"
|
| 39 |
},
|
| 40 |
"THIRD_PARTY_NOTICES.md": {
|
| 41 |
+
"bytes": 4218,
|
| 42 |
+
"sha256": "c358cf50b20dd1d61e2b4418cb47816ab623e4fc249f101dce4a603da7f4b519"
|
| 43 |
},
|
| 44 |
"TRAINING_AND_REPRODUCIBILITY.md": {
|
| 45 |
+
"bytes": 4237,
|
| 46 |
+
"sha256": "7acda8fa1f07a7497fe7354bfeeb709a5f32660fe12815a220d920fdb2bec4a3"
|
| 47 |
},
|
| 48 |
"config.json": {
|
| 49 |
+
"bytes": 923,
|
| 50 |
+
"sha256": "18f8ecd60078be202da13c2345f1b498938800c8ab969f2b49b68237a0f7e977"
|
| 51 |
},
|
| 52 |
"configs/pretrain_20261022_121457_b3d5f398/corpus.json": {
|
| 53 |
"bytes": 1356,
|
|
|
|
| 126 |
"sha256": "ae708eca3c09d5a27eab3fddbe4106d0658787beebc9393de57f75234480a284"
|
| 127 |
},
|
| 128 |
"docs/DATA.md": {
|
| 129 |
+
"bytes": 6927,
|
| 130 |
+
"sha256": "c64a25558379ff1a40b7d2656f9e5980a09ae318ddc29b9564af477551170447"
|
| 131 |
},
|
| 132 |
"docs/EVALUATION.md": {
|
| 133 |
+
"bytes": 8948,
|
| 134 |
+
"sha256": "47778005e7954582eac93ee71186031671cf6e52256f9ee3805a01ce1f08a171"
|
| 135 |
},
|
| 136 |
"docs/EXAMPLES.md": {
|
| 137 |
"bytes": 1117,
|
| 138 |
"sha256": "7a818c365f1c047b349479829396ba5cd841ca8d2f04fa14fc1489083a1fd4b3"
|
| 139 |
},
|
| 140 |
"docs/TECHNICAL_REPORT.md": {
|
| 141 |
+
"bytes": 13527,
|
| 142 |
+
"sha256": "db45336e0e82a33a2b10bc623605f98f25179780f762f0a5d72e8fe9c7a404c4"
|
| 143 |
},
|
| 144 |
"docs/figures/benchmark-comparison.png": {
|
| 145 |
+
"bytes": 127647,
|
| 146 |
+
"sha256": "3bef7d63121782ee43360da609b56e77e012f7cacde2352fdb2e7403ffd4067a"
|
| 147 |
},
|
| 148 |
"docs/figures/benchmark-comparison.svg": {
|
| 149 |
+
"bytes": 92644,
|
| 150 |
+
"sha256": "ff0594356bf29dd00ba163a5506a4bd009b2537ef6281cd0e316932bde68c12c"
|
| 151 |
},
|
| 152 |
"docs/figures/chat-diagnostics.png": {
|
| 153 |
+
"bytes": 110030,
|
| 154 |
+
"sha256": "57735fe8795f07653186ae0ad0f42f9b6864cd427eb8c29d8f2091a328c7bb6f"
|
| 155 |
},
|
| 156 |
"docs/figures/chat-diagnostics.svg": {
|
| 157 |
+
"bytes": 65792,
|
| 158 |
+
"sha256": "a0696ec0291e8321785f8bbe44b4844c7fe71a1cc79c76de7be28e1c8ec05250"
|
| 159 |
},
|
| 160 |
"docs/figures/cooldown-detail.png": {
|
| 161 |
+
"bytes": 185010,
|
| 162 |
+
"sha256": "1ca6d5a062a5cfafbeecd4f76b6853ad8ca2a457b7c63b95a8ad8567be47ec37"
|
| 163 |
},
|
| 164 |
"docs/figures/cooldown-detail.svg": {
|
| 165 |
+
"bytes": 174344,
|
| 166 |
+
"sha256": "5ba56b1227c230fb6e9ffb30933a1e05a11d1aac690908bf47804e8f7c98dae3"
|
| 167 |
},
|
| 168 |
"docs/figures/ifeval-comparison.png": {
|
| 169 |
+
"bytes": 110881,
|
| 170 |
+
"sha256": "0c987246e68b7c7c908cececc944c8c9d4cb3b986998a14c4a0b6f16a64cf61c"
|
| 171 |
},
|
| 172 |
"docs/figures/ifeval-comparison.svg": {
|
| 173 |
+
"bytes": 71454,
|
| 174 |
+
"sha256": "e497628812a14fbc15e5381096e8591ec82898c00e20bb1d23ae1dd71dcdae5d"
|
| 175 |
},
|
| 176 |
"docs/figures/pretraining-overview.png": {
|
| 177 |
+
"bytes": 157875,
|
| 178 |
+
"sha256": "ab172241620cbd05725620e8709bacf3828812e8c2b79db6e8062a26c8e45bf0"
|
| 179 |
},
|
| 180 |
"docs/figures/pretraining-overview.svg": {
|
| 181 |
+
"bytes": 360303,
|
| 182 |
+
"sha256": "8f01a490f4c89876ac1d1177dd9e3c12c7ee833792d001663cbd1e99fcea00eb"
|
| 183 |
},
|
| 184 |
"evaluation/README.md": {
|
| 185 |
"bytes": 1816,
|
|
|
|
| 250 |
"sha256": "e8f5f334901ee37218c0cf0273cc064af24ff00fb6535ca4c5f2f35843e3b410"
|
| 251 |
},
|
| 252 |
"inference_native/__init__.py": {
|
| 253 |
+
"bytes": 55,
|
| 254 |
+
"sha256": "fc888ba4e501527b55c6551b91e7ad2103ea3ab291834c151d73f9ca15726c44"
|
| 255 |
},
|
| 256 |
"inference_native/chat_format.py": {
|
| 257 |
"bytes": 4129,
|
|
|
|
| 278 |
"sha256": "1e3598d6bdf2b16b4b00b3504cf0fd3c3484b69ab6efe97d3f790c437e85c32f"
|
| 279 |
},
|
| 280 |
"provenance/LINEAGE.json": {
|
| 281 |
+
"bytes": 5768,
|
| 282 |
+
"sha256": "4cc5fa3cfd4090a3ec55ef417abb7c771fdd4bdd205ad1ae0ab7f4f28aa03ea7"
|
| 283 |
},
|
| 284 |
"provenance/PRETRAINING_PHASES.json": {
|
| 285 |
"bytes": 41733,
|
|
|
|
| 514 |
"sha256": "cac0e352b21365bfb4fef9f5c265aa36426552bb47851be5fdb5ff585648f69a"
|
| 515 |
},
|
| 516 |
"scripts/plot_results.py": {
|
| 517 |
+
"bytes": 6225,
|
| 518 |
+
"sha256": "648a6478bf70b5fc2732b8f40b35bcc069e099336c52cfd935696218b3e5524b"
|
| 519 |
},
|
| 520 |
"scripts/plot_training.py": {
|
| 521 |
+
"bytes": 2874,
|
| 522 |
+
"sha256": "a96bfb8538756bea2ca76063715fef2b7055f36026833e7b4e4c00f6d6087074"
|
| 523 |
},
|
| 524 |
"scripts/verify_inference.py": {
|
| 525 |
"bytes": 4559,
|
| 526 |
"sha256": "002b4e552d3d625f96d77d90a0e3195e53e5c4358f64fe3a042792d08b232cd3"
|
| 527 |
},
|
| 528 |
"scripts/verify_package.py": {
|
| 529 |
+
"bytes": 1768,
|
| 530 |
+
"sha256": "d3070fc2a2cc3bfd0ccc6e406ffb8831ae977b8319bc0a997642464693e7dbbe"
|
| 531 |
},
|
| 532 |
"tokenizer/corpus.json": {
|
| 533 |
"bytes": 1549,
|
README.md
CHANGED
|
@@ -24,15 +24,17 @@ datasets:
|
|
| 24 |
- allenai/sciq
|
| 25 |
---
|
| 26 |
|
| 27 |
-
# Solstice-1
|
| 28 |
|
| 29 |
**A 175M-parameter English chat model, trained from scratch on five Macs.**
|
| 30 |
|
| 31 |
William Zebrowski 路 OpenSML 路 Apache-2.0 路 Research preview
|
| 32 |
|
| 33 |
-
[Weights](https://huggingface.co/OpenSML/Solstice-1) 路 [Source](https://github.com/williamzebrowskI/Solstice-1) 路 [Run guide](RUN_GUIDE.md) 路 [Technical report](docs/TECHNICAL_REPORT.md) 路 [Evaluation](docs/EVALUATION.md) 路 [Data](docs/DATA.md)
|
| 34 |
|
| 35 |
-
Solstice-1 is the first model in the Solstice family. It combines **42.001B pretraining tokens** with three stages of adapter fine-tuning for short conversations, reading comprehension and instruction following. The release includes merged FP32 weights, the original tokenizer, an MLX runner, training records and benchmark outputs.
|
|
|
|
|
|
|
| 36 |
|
| 37 |
## Model specification
|
| 38 |
|
|
@@ -47,7 +49,7 @@ Solstice-1 is the first model in the Solstice family. It combines **42.001B pret
|
|
| 47 |
| Pretraining | 42,001,367,040 processed token positions |
|
| 48 |
| Release | Merged FP32 safetensors; native MLX on Apple Silicon |
|
| 49 |
|
| 50 |
-
Solstice-1 has a separate weight lineage from OpenSML-150M and OpenSML-JSONLite-175M. Reusing the tokenizer does not make those models its weight ancestors.
|
| 51 |
|
| 52 |
## Measured results
|
| 53 |
|
|
@@ -57,7 +59,7 @@ Zero-shot likelihood scoring on the complete ARC Easy, ARC Challenge, PIQA and H
|
|
| 57 |
|
| 58 |
<table>
|
| 59 |
<tbody>
|
| 60 |
-
<tr><td role="columnheader" align="center"><strong>Benchmark</strong></td><td role="columnheader" align="center"><strong>Solstice-1</strong><br><sub><em>acc / acc_norm</em></sub></td><td role="columnheader" align="center"><strong>OpenSML-150M</strong><br><sub><em>acc / acc_norm</em></sub></td></tr>
|
| 61 |
<tr><td>ARC Easy</td><td><strong>62.58</strong> / <strong>62.33</strong></td><td>56.65 / 55.43</td></tr>
|
| 62 |
<tr><td>ARC Challenge</td><td><strong>29.35</strong> / <strong>30.89</strong></td><td>26.02 / 29.52</td></tr>
|
| 63 |
<tr><td>PIQA</td><td><strong>66.49</strong> / <strong>66.21</strong></td><td>64.53 / 64.09</td></tr>
|
|
@@ -71,7 +73,7 @@ Values are percentages, shown as **acc / acc_norm**. Bold marks the higher score
|
|
| 71 |
|
| 72 |
IFEval checks whether generated responses follow explicit instructions. It measures accuracy, with separate strict and loose scoring at the prompt and instruction levels. **It has no `acc_norm` metric.**
|
| 73 |
|
| 74 |
-
| IFEval accuracy | Solstice-1 | OpenSML-150M |
|
| 75 |
| --- | --- | --- |
|
| 76 |
| Strict prompts | **17.93%** | 15.16% |
|
| 77 |
| Strict instructions | **30.10%** | 25.30% |
|
|
@@ -88,7 +90,7 @@ Both models use FP32 choice scoring, but implementations and chat formats differ
|
|
| 88 |
|
| 89 |
A separate authored chat panel measures short-answer matching and response behavior. These automatic checks are narrower than human judgments of conversation quality.
|
| 90 |
|
| 91 |
-
| Measure | Solstice-1 |
|
| 92 |
| --- | ---: |
|
| 93 |
| Answer accuracy, objective cases | 59.44% |
|
| 94 |
| Natural stopping, all responses | 99.48% |
|
|
@@ -158,8 +160,8 @@ Use Python 3.11 on Apple Silicon. The public model package can be downloaded wit
|
|
| 158 |
python3.11 -m venv .venv
|
| 159 |
source .venv/bin/activate
|
| 160 |
python -m pip install huggingface_hub
|
| 161 |
-
hf download OpenSML/Solstice-1 --local-dir Solstice-1
|
| 162 |
-
cd Solstice-1
|
| 163 |
python -m pip install -r requirements.txt
|
| 164 |
python inference.py --prompt "What is the main function of the heart?" --max-new-tokens 96
|
| 165 |
python inference.py --interactive
|
|
@@ -178,7 +180,7 @@ The supplied runner is the supported inference path. Transformers AutoModel, MLX
|
|
| 178 |
- **Scope:** reliable coding, SQL, JSON extraction, tool use, multilingual use and long conversations are not established capabilities.
|
| 179 |
- **Evaluation coverage:** public and historical tests informed earlier research. The reserved test shares authored task families with training. No comprehensive semantic contamination audit, independent human study or safety assessment is claimed.
|
| 180 |
|
| 181 |
-
Solstice-1 is a research release for short English text and chat. It is not a dependable source of facts or a validated system for consequential decisions.
|
| 182 |
|
| 183 |
## Files, reproducibility and license
|
| 184 |
|
|
@@ -186,4 +188,4 @@ Solstice-1 is a research release for short English text and chat. It is not a de
|
|
| 186 |
|
| 187 |
The repository includes weights on Hugging Face, tokenizer files, inference code, sanitized training snapshots, frozen recipes, selected-checkpoint hashes, curves and saved evaluation outputs. Raw corpora, optimizer binaries and all ancestor weights are not bundled. The training archive documents the run; it is not a verified one-command retraining package.
|
| 188 |
|
| 189 |
-
Weights and author-controlled code use [Apache-2.0](https://huggingface.co/OpenSML/Solstice-1/blob/main/LICENSE). Upstream datasets and third-party code retain their own terms. [Source notice](SOURCE_NOTICE.md) 路 [Third-party notices](THIRD_PARTY_NOTICES.md) 路 [Citation](CITATION.cff)
|
|
|
|
| 24 |
- allenai/sciq
|
| 25 |
---
|
| 26 |
|
| 27 |
+
# Solstice-1-175M
|
| 28 |
|
| 29 |
**A 175M-parameter English chat model, trained from scratch on five Macs.**
|
| 30 |
|
| 31 |
William Zebrowski 路 OpenSML 路 Apache-2.0 路 Research preview
|
| 32 |
|
| 33 |
+
[Weights](https://huggingface.co/OpenSML/Solstice-1-175M) 路 [Source](https://github.com/williamzebrowskI/Solstice-1-175M) 路 [Run guide](RUN_GUIDE.md) 路 [Technical report](docs/TECHNICAL_REPORT.md) 路 [Evaluation](docs/EVALUATION.md) 路 [Data](docs/DATA.md)
|
| 34 |
|
| 35 |
+
Solstice-1-175M is the first model in the Solstice family. It combines **42.001B pretraining tokens** with three stages of adapter fine-tuning for short conversations, reading comprehension and instruction following. The release includes merged FP32 weights, the original tokenizer, an MLX runner, training records and benchmark outputs.
|
| 36 |
+
|
| 37 |
+
Previously published as `OpenSML/Solstice-1`. This naming update leaves the weights, tokenizer, and recorded evaluation results unchanged; historical records retain their original identifiers.
|
| 38 |
|
| 39 |
## Model specification
|
| 40 |
|
|
|
|
| 49 |
| Pretraining | 42,001,367,040 processed token positions |
|
| 50 |
| Release | Merged FP32 safetensors; native MLX on Apple Silicon |
|
| 51 |
|
| 52 |
+
Solstice-1-175M has a separate weight lineage from OpenSML-150M and OpenSML-JSONLite-175M. Reusing the tokenizer does not make those models its weight ancestors.
|
| 53 |
|
| 54 |
## Measured results
|
| 55 |
|
|
|
|
| 59 |
|
| 60 |
<table>
|
| 61 |
<tbody>
|
| 62 |
+
<tr><td role="columnheader" align="center"><strong>Benchmark</strong></td><td role="columnheader" align="center"><strong>Solstice-1-175M</strong><br><sub><em>acc / acc_norm</em></sub></td><td role="columnheader" align="center"><strong>OpenSML-150M</strong><br><sub><em>acc / acc_norm</em></sub></td></tr>
|
| 63 |
<tr><td>ARC Easy</td><td><strong>62.58</strong> / <strong>62.33</strong></td><td>56.65 / 55.43</td></tr>
|
| 64 |
<tr><td>ARC Challenge</td><td><strong>29.35</strong> / <strong>30.89</strong></td><td>26.02 / 29.52</td></tr>
|
| 65 |
<tr><td>PIQA</td><td><strong>66.49</strong> / <strong>66.21</strong></td><td>64.53 / 64.09</td></tr>
|
|
|
|
| 73 |
|
| 74 |
IFEval checks whether generated responses follow explicit instructions. It measures accuracy, with separate strict and loose scoring at the prompt and instruction levels. **It has no `acc_norm` metric.**
|
| 75 |
|
| 76 |
+
| IFEval accuracy | Solstice-1-175M | OpenSML-150M |
|
| 77 |
| --- | --- | --- |
|
| 78 |
| Strict prompts | **17.93%** | 15.16% |
|
| 79 |
| Strict instructions | **30.10%** | 25.30% |
|
|
|
|
| 90 |
|
| 91 |
A separate authored chat panel measures short-answer matching and response behavior. These automatic checks are narrower than human judgments of conversation quality.
|
| 92 |
|
| 93 |
+
| Measure | Solstice-1-175M |
|
| 94 |
| --- | ---: |
|
| 95 |
| Answer accuracy, objective cases | 59.44% |
|
| 96 |
| Natural stopping, all responses | 99.48% |
|
|
|
|
| 160 |
python3.11 -m venv .venv
|
| 161 |
source .venv/bin/activate
|
| 162 |
python -m pip install huggingface_hub
|
| 163 |
+
hf download OpenSML/Solstice-1-175M --local-dir Solstice-1-175M
|
| 164 |
+
cd Solstice-1-175M
|
| 165 |
python -m pip install -r requirements.txt
|
| 166 |
python inference.py --prompt "What is the main function of the heart?" --max-new-tokens 96
|
| 167 |
python inference.py --interactive
|
|
|
|
| 180 |
- **Scope:** reliable coding, SQL, JSON extraction, tool use, multilingual use and long conversations are not established capabilities.
|
| 181 |
- **Evaluation coverage:** public and historical tests informed earlier research. The reserved test shares authored task families with training. No comprehensive semantic contamination audit, independent human study or safety assessment is claimed.
|
| 182 |
|
| 183 |
+
Solstice-1-175M is a research release for short English text and chat. It is not a dependable source of facts or a validated system for consequential decisions.
|
| 184 |
|
| 185 |
## Files, reproducibility and license
|
| 186 |
|
|
|
|
| 188 |
|
| 189 |
The repository includes weights on Hugging Face, tokenizer files, inference code, sanitized training snapshots, frozen recipes, selected-checkpoint hashes, curves and saved evaluation outputs. Raw corpora, optimizer binaries and all ancestor weights are not bundled. The training archive documents the run; it is not a verified one-command retraining package.
|
| 190 |
|
| 191 |
+
Weights and author-controlled code use [Apache-2.0](https://huggingface.co/OpenSML/Solstice-1-175M/blob/main/LICENSE). Upstream datasets and third-party code retain their own terms. [Source notice](SOURCE_NOTICE.md) 路 [Third-party notices](THIRD_PARTY_NOTICES.md) 路 [Citation](CITATION.cff)
|
REPOSITORIES.json
CHANGED
|
@@ -1,17 +1,17 @@
|
|
| 1 |
{
|
| 2 |
-
"model": "Solstice-1",
|
| 3 |
"organization": "OpenSML",
|
| 4 |
"parameter_label": "175M",
|
| 5 |
"license": "Apache-2.0",
|
| 6 |
"release_status": "public research release",
|
| 7 |
"repositories": {
|
| 8 |
"github": {
|
| 9 |
-
"url": "https://github.com/williamzebrowskI/Solstice-1",
|
| 10 |
"visibility": "public",
|
| 11 |
"role": "Native inference code, documentation, training provenance, evaluation evidence"
|
| 12 |
},
|
| 13 |
"huggingface": {
|
| 14 |
-
"url": "https://huggingface.co/OpenSML/Solstice-1",
|
| 15 |
"visibility": "public",
|
| 16 |
"role": "Complete runnable package, including merged FP32 weights and tokenizer"
|
| 17 |
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"model": "Solstice-1-175M",
|
| 3 |
"organization": "OpenSML",
|
| 4 |
"parameter_label": "175M",
|
| 5 |
"license": "Apache-2.0",
|
| 6 |
"release_status": "public research release",
|
| 7 |
"repositories": {
|
| 8 |
"github": {
|
| 9 |
+
"url": "https://github.com/williamzebrowskI/Solstice-1-175M",
|
| 10 |
"visibility": "public",
|
| 11 |
"role": "Native inference code, documentation, training provenance, evaluation evidence"
|
| 12 |
},
|
| 13 |
"huggingface": {
|
| 14 |
+
"url": "https://huggingface.co/OpenSML/Solstice-1-175M",
|
| 15 |
"visibility": "public",
|
| 16 |
"role": "Complete runnable package, including merged FP32 weights and tokenizer"
|
| 17 |
}
|
RUN_GUIDE.md
CHANGED
|
@@ -1,4 +1,4 @@
|
|
| 1 |
-
# Run Solstice-1
|
| 2 |
|
| 3 |
Use Python 3.11 on Apple Silicon macOS with the pinned requirements. The tested public MLX wheel is 0.32.3; it requires a compatible macOS version. See PACKAGING_VERIFICATION.json for the actual tested system. Inference is offline after downloading the package and installing dependencies.
|
| 4 |
|
|
@@ -8,8 +8,8 @@ The model package is public. No Hugging Face sign-in is required to download it.
|
|
| 8 |
python3.11 -m venv .venv
|
| 9 |
source .venv/bin/activate
|
| 10 |
python -m pip install huggingface_hub
|
| 11 |
-
hf download OpenSML/Solstice-1 --local-dir Solstice-1
|
| 12 |
-
cd Solstice-1
|
| 13 |
python -m pip install -r requirements.txt
|
| 14 |
python scripts/verify_package.py
|
| 15 |
python inference.py --prompt "Name the largest continent." --max-new-tokens 64
|
|
|
|
| 1 |
+
# Run Solstice-1-175M
|
| 2 |
|
| 3 |
Use Python 3.11 on Apple Silicon macOS with the pinned requirements. The tested public MLX wheel is 0.32.3; it requires a compatible macOS version. See PACKAGING_VERIFICATION.json for the actual tested system. Inference is offline after downloading the package and installing dependencies.
|
| 4 |
|
|
|
|
| 8 |
python3.11 -m venv .venv
|
| 9 |
source .venv/bin/activate
|
| 10 |
python -m pip install huggingface_hub
|
| 11 |
+
hf download OpenSML/Solstice-1-175M --local-dir Solstice-1-175M
|
| 12 |
+
cd Solstice-1-175M
|
| 13 |
python -m pip install -r requirements.txt
|
| 14 |
python scripts/verify_package.py
|
| 15 |
python inference.py --prompt "Name the largest continent." --max-new-tokens 64
|
SHA256SUMS
CHANGED
|
@@ -1,15 +1,16 @@
|
|
| 1 |
260d3301bac863ab1b0a29739ff5f5fee1a5ec2fa2dccdd03353f07bcf6292c2 .gitignore
|
| 2 |
f09ed498cc42b482dae88d9fb5026bf74e1d8b9187f9ea301aeda0db76d6c859 BACKUP_AND_CLEANUP.md
|
| 3 |
-
|
|
|
|
| 4 |
cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30 LICENSE
|
| 5 |
0e6865f1a81cfd69ca48e6987658854ce0c0d23e91d72ae545393cd5fa032375 PACKAGING_VERIFICATION.json
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632 configs/pretrain_20261022_121457_b3d5f398/corpus.json
|
| 14 |
ddf14c4cef8a4fb66392a27e87f8c5a807a88da35fe61599f252c9eba91bec29 configs/pretrain_20261022_121457_b3d5f398/recipe.json
|
| 15 |
c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632 configs/pretrain_20261024_111140_5f596b7a/corpus.json
|
|
@@ -29,20 +30,20 @@ c70c4491d8865d0db8af21d555c309dcfcbaa346760f0668e82a8c0ce54c8d6f configs/pretra
|
|
| 29 |
fb204be1d9e72eb1cae140bfdb03dbcadc1843c47b1894a731396ceeedb9a22f configs/sft1/plan.json
|
| 30 |
51d66108416988bad1015d7b52cf63db0571f87f202b2b18c7e98e67a29ec62b configs/sft2/plan.json
|
| 31 |
ae708eca3c09d5a27eab3fddbe4106d0658787beebc9393de57f75234480a284 configs/sft3/plan.json
|
| 32 |
-
|
| 33 |
-
|
| 34 |
7a818c365f1c047b349479829396ba5cd841ca8d2f04fa14fc1489083a1fd4b3 docs/EXAMPLES.md
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
9e15b1b93bbc1160f3988a7956eee88b2467e5d826f097df35401562c116b90d evaluation/README.md
|
| 47 |
f3835c8cb0bdf580e4142a57ac3724c474dcc1d44aa01dbbd805c989a587c7c7 evaluation/benchmark_best.py
|
| 48 |
10ae9d2c49236c377d6e9952674b14289f5f92c310c29cbe592a40c98a14c752 evaluation/benchmark_ifeval.py
|
|
@@ -60,14 +61,14 @@ a73797261eee5bf447e279d82a2b700b1bdd3cb1193412dbab1270a85832bc6b evaluation/ven
|
|
| 60 |
5e3542d6369138730a11db9b7793fe1a722021b66ea8ad32b2e93d49ade91555 evaluation/vendor/ifeval/prepared_runtime.json
|
| 61 |
d1ccd2965b9b8752d51b1e814faa5c4879e0c630047fb01c4fa50dca75527d70 evaluation/vendor/ifeval/source.json
|
| 62 |
e8f5f334901ee37218c0cf0273cc064af24ff00fb6535ca4c5f2f35843e3b410 inference.py
|
| 63 |
-
|
| 64 |
5c558b0cfd4d2061af144c3f295b7bec31d92cb4b0fb605f702972c373e2d1e8 inference_native/chat_format.py
|
| 65 |
d9bdeeb4b8ea60dacb9af1cafa1dfbd295bfb384c2070d282848ec7e5eb42168 inference_native/common.py
|
| 66 |
1eee6f8afe78767ff6b9c604ccf2328c98b24bc94dbad6291a0ad699dd7345fe inference_native/model.py
|
| 67 |
254e1f5d60cc4498a9f1209955752db28b91839f48d3121d674f3f5ac42e52e5 inference_native/tokenization.py
|
| 68 |
7b2d8777040f05d8e0849acac4e45013b2de3d72145c363cd22825e80115c515 model.safetensors
|
| 69 |
1e3598d6bdf2b16b4b00b3504cf0fd3c3484b69ab6efe97d3f790c437e85c32f provenance/INFERENCE_EXPORT.json
|
| 70 |
-
|
| 71 |
0db456c4bf1a5e46f831195346626b73d953d412b23c47bc6b27eaa230babf42 provenance/PRETRAINING_PHASES.json
|
| 72 |
d8aaf1ef62597ac384b6f7c9510b24c2bbb63470150c3835266203ee6a567fd8 provenance/SELECTED_PREFIX.json
|
| 73 |
834edc532edfe9a897b60a68046006db6700bdd21de72f055199eb827aef401a provenance/TENSOR_INVENTORY.json
|
|
@@ -126,10 +127,10 @@ da83222ee38d33eea6c7062c28fd310e61da89a9a0315dc62559c4a9181759d9 results/search
|
|
| 126 |
a478bc790345129fa960060490d4c2148351ccfc17ac433ec06364969787a8fe results/selection/design.json
|
| 127 |
0567b916bdf27de7f6eafe8c61da7fb3686a69e08361ac6cb5645ac567cf8f69 results/selection/selection_frozen.json
|
| 128 |
cac0e352b21365bfb4fef9f5c265aa36426552bb47851be5fdb5ff585648f69a scripts/materialize_training_snapshot.py
|
| 129 |
-
|
| 130 |
-
|
| 131 |
002b4e552d3d625f96d77d90a0e3195e53e5c4358f64fe3a042792d08b232cd3 scripts/verify_inference.py
|
| 132 |
-
|
| 133 |
e4c47b7b6aba08f948de7c2c7cb94c58dbd09dde2ded6ac185c60856ee686d27 tokenizer/corpus.json
|
| 134 |
1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb tokenizer/manifest.json
|
| 135 |
e871f690eabebf82da173f0e406f38b79a10cd7d2647bb3592c3fc35ae9a2f16 tokenizer/report.json
|
|
@@ -473,4 +474,3 @@ ff056da7218dfbcae9d5c30e1a7770739683866bff398eebc05f3cfd686b1957 training/sourc
|
|
| 473 |
ff84dc0dfacceb506e54b6af73c1a76d65fcbf9cf75c7d96fc4489b8ab4f79e6 training/source_blobs/ff84dc0dfacceb506e54b6af73c1a76d65fcbf9cf75c7d96fc4489b8ab4f79e6.py
|
| 474 |
ff9cd7e99a7b66c901137a7f61c97c9ecbd3e5effec879cb82a5c68e9ced35f7 training/source_blobs/ff9cd7e99a7b66c901137a7f61c97c9ecbd3e5effec879cb82a5c68e9ced35f7.py
|
| 475 |
ffc3d395e98ab0f55c9d38060d65ff8dcdae4b954a25aafb40c14f513d8067ba training/source_blobs/ffc3d395e98ab0f55c9d38060d65ff8dcdae4b954a25aafb40c14f513d8067ba.py
|
| 476 |
-
4831e3256d28b7a04a99b60dc1fcceb8f1e0f29225b97931b843e30985f37f6c FILE_MANIFEST.json
|
|
|
|
| 1 |
260d3301bac863ab1b0a29739ff5f5fee1a5ec2fa2dccdd03353f07bcf6292c2 .gitignore
|
| 2 |
f09ed498cc42b482dae88d9fb5026bf74e1d8b9187f9ea301aeda0db76d6c859 BACKUP_AND_CLEANUP.md
|
| 3 |
+
097c833a247c878ebc6e4b070e861e1637cf90bd25b5206fe7f3dc2bbcd0e3a4 CITATION.cff
|
| 4 |
+
9e162d73747917683a7bf1f95fe626ac08bc16bbdea4e2022485256f94e6fd69 FILE_MANIFEST.json
|
| 5 |
cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30 LICENSE
|
| 6 |
0e6865f1a81cfd69ca48e6987658854ce0c0d23e91d72ae545393cd5fa032375 PACKAGING_VERIFICATION.json
|
| 7 |
+
dc72b70e5db4fdf9e7a5b8307f7801c6de12e4ba45555349bd0413438ad5333a README.md
|
| 8 |
+
572c3931bd1b8278208873863148b2dd99cb9882accc431fa52143d93dac0d5d REPOSITORIES.json
|
| 9 |
+
09554010710c463910e392079ab69ed5de4314321afb7a260a6ff4ebe9160ce5 RUN_GUIDE.md
|
| 10 |
+
819ca2766a86d1ff4e8a5265a9d03d4ddc315a9256e2c008626c9e0df462489b SOURCE_NOTICE.md
|
| 11 |
+
c358cf50b20dd1d61e2b4418cb47816ab623e4fc249f101dce4a603da7f4b519 THIRD_PARTY_NOTICES.md
|
| 12 |
+
7acda8fa1f07a7497fe7354bfeeb709a5f32660fe12815a220d920fdb2bec4a3 TRAINING_AND_REPRODUCIBILITY.md
|
| 13 |
+
18f8ecd60078be202da13c2345f1b498938800c8ab969f2b49b68237a0f7e977 config.json
|
| 14 |
c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632 configs/pretrain_20261022_121457_b3d5f398/corpus.json
|
| 15 |
ddf14c4cef8a4fb66392a27e87f8c5a807a88da35fe61599f252c9eba91bec29 configs/pretrain_20261022_121457_b3d5f398/recipe.json
|
| 16 |
c35100988b38c2e4bdb141e2bee37e73e731cc9240891360f6a0aba310be0632 configs/pretrain_20261024_111140_5f596b7a/corpus.json
|
|
|
|
| 30 |
fb204be1d9e72eb1cae140bfdb03dbcadc1843c47b1894a731396ceeedb9a22f configs/sft1/plan.json
|
| 31 |
51d66108416988bad1015d7b52cf63db0571f87f202b2b18c7e98e67a29ec62b configs/sft2/plan.json
|
| 32 |
ae708eca3c09d5a27eab3fddbe4106d0658787beebc9393de57f75234480a284 configs/sft3/plan.json
|
| 33 |
+
c64a25558379ff1a40b7d2656f9e5980a09ae318ddc29b9564af477551170447 docs/DATA.md
|
| 34 |
+
47778005e7954582eac93ee71186031671cf6e52256f9ee3805a01ce1f08a171 docs/EVALUATION.md
|
| 35 |
7a818c365f1c047b349479829396ba5cd841ca8d2f04fa14fc1489083a1fd4b3 docs/EXAMPLES.md
|
| 36 |
+
db45336e0e82a33a2b10bc623605f98f25179780f762f0a5d72e8fe9c7a404c4 docs/TECHNICAL_REPORT.md
|
| 37 |
+
3bef7d63121782ee43360da609b56e77e012f7cacde2352fdb2e7403ffd4067a docs/figures/benchmark-comparison.png
|
| 38 |
+
ff0594356bf29dd00ba163a5506a4bd009b2537ef6281cd0e316932bde68c12c docs/figures/benchmark-comparison.svg
|
| 39 |
+
57735fe8795f07653186ae0ad0f42f9b6864cd427eb8c29d8f2091a328c7bb6f docs/figures/chat-diagnostics.png
|
| 40 |
+
a0696ec0291e8321785f8bbe44b4844c7fe71a1cc79c76de7be28e1c8ec05250 docs/figures/chat-diagnostics.svg
|
| 41 |
+
1ca6d5a062a5cfafbeecd4f76b6853ad8ca2a457b7c63b95a8ad8567be47ec37 docs/figures/cooldown-detail.png
|
| 42 |
+
5ba56b1227c230fb6e9ffb30933a1e05a11d1aac690908bf47804e8f7c98dae3 docs/figures/cooldown-detail.svg
|
| 43 |
+
0c987246e68b7c7c908cececc944c8c9d4cb3b986998a14c4a0b6f16a64cf61c docs/figures/ifeval-comparison.png
|
| 44 |
+
e497628812a14fbc15e5381096e8591ec82898c00e20bb1d23ae1dd71dcdae5d docs/figures/ifeval-comparison.svg
|
| 45 |
+
ab172241620cbd05725620e8709bacf3828812e8c2b79db6e8062a26c8e45bf0 docs/figures/pretraining-overview.png
|
| 46 |
+
8f01a490f4c89876ac1d1177dd9e3c12c7ee833792d001663cbd1e99fcea00eb docs/figures/pretraining-overview.svg
|
| 47 |
9e15b1b93bbc1160f3988a7956eee88b2467e5d826f097df35401562c116b90d evaluation/README.md
|
| 48 |
f3835c8cb0bdf580e4142a57ac3724c474dcc1d44aa01dbbd805c989a587c7c7 evaluation/benchmark_best.py
|
| 49 |
10ae9d2c49236c377d6e9952674b14289f5f92c310c29cbe592a40c98a14c752 evaluation/benchmark_ifeval.py
|
|
|
|
| 61 |
5e3542d6369138730a11db9b7793fe1a722021b66ea8ad32b2e93d49ade91555 evaluation/vendor/ifeval/prepared_runtime.json
|
| 62 |
d1ccd2965b9b8752d51b1e814faa5c4879e0c630047fb01c4fa50dca75527d70 evaluation/vendor/ifeval/source.json
|
| 63 |
e8f5f334901ee37218c0cf0273cc064af24ff00fb6535ca4c5f2f35843e3b410 inference.py
|
| 64 |
+
fc888ba4e501527b55c6551b91e7ad2103ea3ab291834c151d73f9ca15726c44 inference_native/__init__.py
|
| 65 |
5c558b0cfd4d2061af144c3f295b7bec31d92cb4b0fb605f702972c373e2d1e8 inference_native/chat_format.py
|
| 66 |
d9bdeeb4b8ea60dacb9af1cafa1dfbd295bfb384c2070d282848ec7e5eb42168 inference_native/common.py
|
| 67 |
1eee6f8afe78767ff6b9c604ccf2328c98b24bc94dbad6291a0ad699dd7345fe inference_native/model.py
|
| 68 |
254e1f5d60cc4498a9f1209955752db28b91839f48d3121d674f3f5ac42e52e5 inference_native/tokenization.py
|
| 69 |
7b2d8777040f05d8e0849acac4e45013b2de3d72145c363cd22825e80115c515 model.safetensors
|
| 70 |
1e3598d6bdf2b16b4b00b3504cf0fd3c3484b69ab6efe97d3f790c437e85c32f provenance/INFERENCE_EXPORT.json
|
| 71 |
+
4cc5fa3cfd4090a3ec55ef417abb7c771fdd4bdd205ad1ae0ab7f4f28aa03ea7 provenance/LINEAGE.json
|
| 72 |
0db456c4bf1a5e46f831195346626b73d953d412b23c47bc6b27eaa230babf42 provenance/PRETRAINING_PHASES.json
|
| 73 |
d8aaf1ef62597ac384b6f7c9510b24c2bbb63470150c3835266203ee6a567fd8 provenance/SELECTED_PREFIX.json
|
| 74 |
834edc532edfe9a897b60a68046006db6700bdd21de72f055199eb827aef401a provenance/TENSOR_INVENTORY.json
|
|
|
|
| 127 |
a478bc790345129fa960060490d4c2148351ccfc17ac433ec06364969787a8fe results/selection/design.json
|
| 128 |
0567b916bdf27de7f6eafe8c61da7fb3686a69e08361ac6cb5645ac567cf8f69 results/selection/selection_frozen.json
|
| 129 |
cac0e352b21365bfb4fef9f5c265aa36426552bb47851be5fdb5ff585648f69a scripts/materialize_training_snapshot.py
|
| 130 |
+
648a6478bf70b5fc2732b8f40b35bcc069e099336c52cfd935696218b3e5524b scripts/plot_results.py
|
| 131 |
+
a96bfb8538756bea2ca76063715fef2b7055f36026833e7b4e4c00f6d6087074 scripts/plot_training.py
|
| 132 |
002b4e552d3d625f96d77d90a0e3195e53e5c4358f64fe3a042792d08b232cd3 scripts/verify_inference.py
|
| 133 |
+
d3070fc2a2cc3bfd0ccc6e406ffb8831ae977b8319bc0a997642464693e7dbbe scripts/verify_package.py
|
| 134 |
e4c47b7b6aba08f948de7c2c7cb94c58dbd09dde2ded6ac185c60856ee686d27 tokenizer/corpus.json
|
| 135 |
1194292c6d906ac19e6f01c6ad2b42825143fad426a22aa8856d903776701ecb tokenizer/manifest.json
|
| 136 |
e871f690eabebf82da173f0e406f38b79a10cd7d2647bb3592c3fc35ae9a2f16 tokenizer/report.json
|
|
|
|
| 474 |
ff84dc0dfacceb506e54b6af73c1a76d65fcbf9cf75c7d96fc4489b8ab4f79e6 training/source_blobs/ff84dc0dfacceb506e54b6af73c1a76d65fcbf9cf75c7d96fc4489b8ab4f79e6.py
|
| 475 |
ff9cd7e99a7b66c901137a7f61c97c9ecbd3e5effec879cb82a5c68e9ced35f7 training/source_blobs/ff9cd7e99a7b66c901137a7f61c97c9ecbd3e5effec879cb82a5c68e9ced35f7.py
|
| 476 |
ffc3d395e98ab0f55c9d38060d65ff8dcdae4b954a25aafb40c14f513d8067ba training/source_blobs/ffc3d395e98ab0f55c9d38060d65ff8dcdae4b954a25aafb40c14f513d8067ba.py
|
|
|
SOURCE_NOTICE.md
CHANGED
|
@@ -2,7 +2,7 @@
|
|
| 2 |
|
| 3 |
Copyright 2026 William Zebrowski.
|
| 4 |
|
| 5 |
-
The selected Solstice-1 model weights, frozen tokenizer artifacts to the extent controlled by the author, and author-controlled project code/documentation are released under Apache License 2.0. See LICENSE. This is the author's explicit release decision for Solstice-1, rather than an inferred license from earlier OpenSML releases.
|
| 6 |
|
| 7 |
Third-party software and training materials retain their own licenses and attribution. Apache-2.0 does not relicense upstream datasets, their underlying documents/code repositories, or separately licensed vendored components. Raw training datasets are not redistributed here. See THIRD_PARTY_NOTICES.md and docs/DATA.md.
|
| 8 |
|
|
|
|
| 2 |
|
| 3 |
Copyright 2026 William Zebrowski.
|
| 4 |
|
| 5 |
+
The selected Solstice-1-175M model weights, frozen tokenizer artifacts to the extent controlled by the author, and author-controlled project code/documentation are released under Apache License 2.0. See LICENSE. This is the author's explicit release decision for Solstice-1-175M, rather than an inferred license from earlier OpenSML releases.
|
| 6 |
|
| 7 |
Third-party software and training materials retain their own licenses and attribution. Apache-2.0 does not relicense upstream datasets, their underlying documents/code repositories, or separately licensed vendored components. Raw training datasets are not redistributed here. See THIRD_PARTY_NOTICES.md and docs/DATA.md.
|
| 8 |
|
THIRD_PARTY_NOTICES.md
CHANGED
|
@@ -2,7 +2,7 @@
|
|
| 2 |
|
| 3 |
## Runtime and source
|
| 4 |
|
| 5 |
-
MLX is an Apple open-source project: [ml-explore/mlx](https://github.com/ml-explore/mlx). NumPy and Hugging Face Tokenizers are dependencies distributed under their own terms. Their packages are installed separately and are not relicensed by Solstice-1.
|
| 6 |
|
| 7 |
Archived training code includes SQLGlot source and its retained license where required by the historical code filter. Snapshot manifests preserve the license file mapping. Google IFEval evaluator source and its Apache-2.0 attribution are preserved with the evaluation archive. Dataset content is not made Apache-2.0 by being used in training or evaluation.
|
| 8 |
|
|
@@ -29,4 +29,4 @@ SciQ is recorded as CC-BY-NC-3.0 and SQuAD v2 as CC-BY-SA-4.0. SmolTalk-family s
|
|
| 29 |
|
| 30 |
## Prior OpenSML references
|
| 31 |
|
| 32 |
-
[OpenSML-150M](https://huggingface.co/OpenSML/OpenSML-150M) and [OpenSML-JSONLite-175M](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M) are earlier releases by the same author. Their weights are not ancestors of Solstice-1. Prior documentation layouts informed this package; their task-specific results are not silently assigned to this model.
|
|
|
|
| 2 |
|
| 3 |
## Runtime and source
|
| 4 |
|
| 5 |
+
MLX is an Apple open-source project: [ml-explore/mlx](https://github.com/ml-explore/mlx). NumPy and Hugging Face Tokenizers are dependencies distributed under their own terms. Their packages are installed separately and are not relicensed by Solstice-1-175M.
|
| 6 |
|
| 7 |
Archived training code includes SQLGlot source and its retained license where required by the historical code filter. Snapshot manifests preserve the license file mapping. Google IFEval evaluator source and its Apache-2.0 attribution are preserved with the evaluation archive. Dataset content is not made Apache-2.0 by being used in training or evaluation.
|
| 8 |
|
|
|
|
| 29 |
|
| 30 |
## Prior OpenSML references
|
| 31 |
|
| 32 |
+
[OpenSML-150M](https://huggingface.co/OpenSML/OpenSML-150M) and [OpenSML-JSONLite-175M](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M) are earlier releases by the same author. Their weights are not ancestors of Solstice-1-175M. Prior documentation layouts informed this package; their task-specific results are not silently assigned to this model.
|
TRAINING_AND_REPRODUCIBILITY.md
CHANGED
|
@@ -33,7 +33,7 @@ The original training runtime used MLX `0.32.3.dev20260920+7241f12`. Packaged in
|
|
| 33 |
|
| 34 |
Absolute workspace/user paths become `${WORKSPACE}` / `${USER_HOME}` in archived configuration and source. Local network addresses are removed where present. Snapshot manifests retain original hashes and packaged hashes separately. Replacing placeholders and satisfying historical file guards is an explicit reproduction prerequisite. Private SSH configurations, network inventories, credentials and raw operational logs are not included.
|
| 35 |
|
| 36 |
-
The release source does not pretend that every imported helper recipe was a model ancestor. Only the phase/configuration and checkpoint lineage records identify selected training exposure. In particular, 167, 1200, 417, previous full-weight SFT attempts and DPO experiments are not additional ancestors of Solstice-1.
|
| 37 |
|
| 38 |
## Evaluations
|
| 39 |
|
|
|
|
| 33 |
|
| 34 |
Absolute workspace/user paths become `${WORKSPACE}` / `${USER_HOME}` in archived configuration and source. Local network addresses are removed where present. Snapshot manifests retain original hashes and packaged hashes separately. Replacing placeholders and satisfying historical file guards is an explicit reproduction prerequisite. Private SSH configurations, network inventories, credentials and raw operational logs are not included.
|
| 35 |
|
| 36 |
+
The release source does not pretend that every imported helper recipe was a model ancestor. Only the phase/configuration and checkpoint lineage records identify selected training exposure. In particular, 167, 1200, 417, previous full-weight SFT attempts and DPO experiments are not additional ancestors of Solstice-1-175M.
|
| 37 |
|
| 38 |
## Evaluations
|
| 39 |
|
config.json
CHANGED
|
@@ -20,8 +20,8 @@
|
|
| 20 |
"qk_norm": true,
|
| 21 |
"vocab_size": 32000
|
| 22 |
},
|
| 23 |
-
"model_id": "OpenSML/Solstice-1",
|
| 24 |
-
"model_name": "Solstice-1",
|
| 25 |
"organization": "OpenSML",
|
| 26 |
"parameter_label": "175M",
|
| 27 |
"parameters": 174202432,
|
|
|
|
| 20 |
"qk_norm": true,
|
| 21 |
"vocab_size": 32000
|
| 22 |
},
|
| 23 |
+
"model_id": "OpenSML/Solstice-1-175M",
|
| 24 |
+
"model_name": "Solstice-1-175M",
|
| 25 |
"organization": "OpenSML",
|
| 26 |
"parameter_label": "175M",
|
| 27 |
"parameters": 174202432,
|
docs/DATA.md
CHANGED
|
@@ -1,4 +1,4 @@
|
|
| 1 |
-
# Solstice-1 data provenance
|
| 2 |
|
| 3 |
## Pretraining
|
| 4 |
|
|
|
|
| 1 |
+
# Solstice-1-175M data provenance
|
| 2 |
|
| 3 |
## Pretraining
|
| 4 |
|
docs/EVALUATION.md
CHANGED
|
@@ -1,4 +1,4 @@
|
|
| 1 |
-
# Solstice-1 evaluation
|
| 2 |
|
| 3 |
All results identify the same release weight hash, `7b2d8777040f05d8e0849acac4e45013b2de3d72145c363cd22825e80115c515`. Public summary: [results/public/summary.json](../results/public/summary.json). The package verification is a separate check of loading/output parity, not a new training result.
|
| 4 |
|
|
@@ -6,7 +6,7 @@ All results identify the same release weight hash, `7b2d8777040f05d8e0849acac4e4
|
|
| 6 |
|
| 7 |
<table>
|
| 8 |
<tbody>
|
| 9 |
-
<tr><td role="columnheader" align="center"><strong>Benchmark</strong></td><td role="columnheader" align="center"><strong>Solstice-1</strong><br><sub><em>acc / acc_norm</em></sub></td><td role="columnheader" align="center"><strong>OpenSML-150M</strong><br><sub><em>acc / acc_norm</em></sub></td><td role="columnheader" align="center"><strong>SmolLM2-135M</strong><br><sub><em>acc / acc_norm</em></sub></td></tr>
|
| 10 |
<tr><td>ARC Easy</td><td><strong>62.58</strong> / <strong>62.33</strong></td><td>56.65 / 55.43</td><td>54.00 / 48.82</td></tr>
|
| 11 |
<tr><td>ARC Challenge</td><td><strong>29.35</strong> / <strong>30.89</strong></td><td>26.02 / 29.52</td><td>25.94 / 27.73</td></tr>
|
| 12 |
<tr><td>PIQA</td><td>66.49 / 66.21</td><td>64.53 / 64.09</td><td><strong>66.70</strong> / <strong>66.76</strong></td></tr>
|
|
@@ -20,7 +20,7 @@ Native MLX evaluates complete ARC Easy/Challenge test splits, PIQA validation an
|
|
| 20 |
|
| 21 |
## Instruction following
|
| 22 |
|
| 23 |
-
| IFEval accuracy | Solstice-1 | OpenSML-150M | SmolLM2-135M |
|
| 24 |
| --- | --- | --- | --- |
|
| 25 |
| Strict prompts | 17.93% | 15.16% | **21.63%** |
|
| 26 |
| Strict instructions | 30.10% | 25.30% | **35.85%** |
|
|
@@ -39,7 +39,7 @@ The independent evaluation pins SmolLM2 to revision `12fd25f77366fa6b3b4b768ec30
|
|
| 39 |
|
| 40 |
## Historical comparison
|
| 41 |
|
| 42 |
-
| Metric | Published 150M | 167 | Parent 2000 | Prior attempt 5 | Solstice-1 |
|
| 43 |
| --- | --- | --- | --- | --- | --- |
|
| 44 |
| ARC Easy acc_norm | 55.43% | 61.49% | **62.33%** | **62.33%** | **62.33%** |
|
| 45 |
| ARC Challenge acc_norm | 29.52% | 30.46% | 30.63% | **31.23%** | 30.89% |
|
|
@@ -56,7 +56,7 @@ JSONLite has different extraction panels, so its 89.54% extraction correctness a
|
|
| 56 |
|
| 57 |
## Authored checks: keep denominators separate
|
| 58 |
|
| 59 |
-
| Panel | Parent 2000 | Prior attempt 5 | Solstice-1 |
|
| 60 |
| --- | ---: | ---: | ---: |
|
| 61 |
| Historical development answers | 197/320 | 204/320 | 200/320 |
|
| 62 |
| Historical final answers | 367/704 | 384/704 | 379/704 |
|
|
|
|
| 1 |
+
# Solstice-1-175M evaluation
|
| 2 |
|
| 3 |
All results identify the same release weight hash, `7b2d8777040f05d8e0849acac4e45013b2de3d72145c363cd22825e80115c515`. Public summary: [results/public/summary.json](../results/public/summary.json). The package verification is a separate check of loading/output parity, not a new training result.
|
| 4 |
|
|
|
|
| 6 |
|
| 7 |
<table>
|
| 8 |
<tbody>
|
| 9 |
+
<tr><td role="columnheader" align="center"><strong>Benchmark</strong></td><td role="columnheader" align="center"><strong>Solstice-1-175M</strong><br><sub><em>acc / acc_norm</em></sub></td><td role="columnheader" align="center"><strong>OpenSML-150M</strong><br><sub><em>acc / acc_norm</em></sub></td><td role="columnheader" align="center"><strong>SmolLM2-135M</strong><br><sub><em>acc / acc_norm</em></sub></td></tr>
|
| 10 |
<tr><td>ARC Easy</td><td><strong>62.58</strong> / <strong>62.33</strong></td><td>56.65 / 55.43</td><td>54.00 / 48.82</td></tr>
|
| 11 |
<tr><td>ARC Challenge</td><td><strong>29.35</strong> / <strong>30.89</strong></td><td>26.02 / 29.52</td><td>25.94 / 27.73</td></tr>
|
| 12 |
<tr><td>PIQA</td><td>66.49 / 66.21</td><td>64.53 / 64.09</td><td><strong>66.70</strong> / <strong>66.76</strong></td></tr>
|
|
|
|
| 20 |
|
| 21 |
## Instruction following
|
| 22 |
|
| 23 |
+
| IFEval accuracy | Solstice-1-175M | OpenSML-150M | SmolLM2-135M |
|
| 24 |
| --- | --- | --- | --- |
|
| 25 |
| Strict prompts | 17.93% | 15.16% | **21.63%** |
|
| 26 |
| Strict instructions | 30.10% | 25.30% | **35.85%** |
|
|
|
|
| 39 |
|
| 40 |
## Historical comparison
|
| 41 |
|
| 42 |
+
| Metric | Published 150M | 167 | Parent 2000 | Prior attempt 5 | Solstice-1-175M |
|
| 43 |
| --- | --- | --- | --- | --- | --- |
|
| 44 |
| ARC Easy acc_norm | 55.43% | 61.49% | **62.33%** | **62.33%** | **62.33%** |
|
| 45 |
| ARC Challenge acc_norm | 29.52% | 30.46% | 30.63% | **31.23%** | 30.89% |
|
|
|
|
| 56 |
|
| 57 |
## Authored checks: keep denominators separate
|
| 58 |
|
| 59 |
+
| Panel | Parent 2000 | Prior attempt 5 | Solstice-1-175M |
|
| 60 |
| --- | ---: | ---: | ---: |
|
| 61 |
| Historical development answers | 197/320 | 204/320 | 200/320 |
|
| 62 |
| Historical final answers | 367/704 | 384/704 | 379/704 |
|
docs/TECHNICAL_REPORT.md
CHANGED
|
@@ -1,10 +1,10 @@
|
|
| 1 |
-
# Solstice-1 technical report
|
| 2 |
|
| 3 |
Author: William Zebrowski 路 Release candidate prepared 2026-11-09 路 Public size designation: **175M**
|
| 4 |
|
| 5 |
## 1. Scope and identity
|
| 6 |
|
| 7 |
-
Solstice-1 is the merged FP32 export of final-stage checkpoint `step_0000128_5937a6def881`. It contains exactly **174,202,432 parameters**. The 175M name is a rounded public designation. This report describes the selected checkpoint rather than all experiments in the workspace.
|
| 8 |
|
| 9 |
Weight SHA-256: `7b2d8777040f05d8e0849acac4e45013b2de3d72145c363cd22825e80115c515`.
|
| 10 |
|
|
@@ -56,7 +56,7 @@ Nominal validation and checkpoint intervals are 100M tokens. Recorded training a
|
|
| 56 |
|
| 57 |

|
| 58 |
|
| 59 |
-
The base checkpoint's saved BF16-path best prose validation loss is 2.8356336173. The FP32 base measurement used by SFT retention checks is 2.8353193145. Solstice-1's FP32 retention loss is 2.8633367460. These precision-specific measurements must not be silently combined. Python and SQL validation losses are separate from prose-best selection. Losses on these datasets cannot be compared directly with unrelated models' losses.
|
| 60 |
|
| 61 |
## 6. Selected fine-tuning lineage
|
| 62 |
|
|
|
|
| 1 |
+
# Solstice-1-175M technical report
|
| 2 |
|
| 3 |
Author: William Zebrowski 路 Release candidate prepared 2026-11-09 路 Public size designation: **175M**
|
| 4 |
|
| 5 |
## 1. Scope and identity
|
| 6 |
|
| 7 |
+
Solstice-1-175M is the merged FP32 export of final-stage checkpoint `step_0000128_5937a6def881`. It contains exactly **174,202,432 parameters**. The 175M name is a rounded public designation. This report describes the selected checkpoint rather than all experiments in the workspace.
|
| 8 |
|
| 9 |
Weight SHA-256: `7b2d8777040f05d8e0849acac4e45013b2de3d72145c363cd22825e80115c515`.
|
| 10 |
|
|
|
|
| 56 |
|
| 57 |

|
| 58 |
|
| 59 |
+
The base checkpoint's saved BF16-path best prose validation loss is 2.8356336173. The FP32 base measurement used by SFT retention checks is 2.8353193145. Solstice-1-175M's FP32 retention loss is 2.8633367460. These precision-specific measurements must not be silently combined. Python and SQL validation losses are separate from prose-best selection. Losses on these datasets cannot be compared directly with unrelated models' losses.
|
| 60 |
|
| 61 |
## 6. Selected fine-tuning lineage
|
| 62 |
|
docs/figures/benchmark-comparison.png
CHANGED
|
Git LFS Details
|
|
Git LFS Details
|
docs/figures/benchmark-comparison.svg
CHANGED
|
|
|
|
docs/figures/chat-diagnostics.png
CHANGED
|
Git LFS Details
|
|
Git LFS Details
|
docs/figures/chat-diagnostics.svg
CHANGED
|
|
|
|
docs/figures/cooldown-detail.png
CHANGED
|
Git LFS Details
|
|
Git LFS Details
|
docs/figures/cooldown-detail.svg
CHANGED
|
|
|
|
docs/figures/ifeval-comparison.png
CHANGED
|
Git LFS Details
|
|
Git LFS Details
|
docs/figures/ifeval-comparison.svg
CHANGED
|
|
|
|
docs/figures/pretraining-overview.png
CHANGED
|
Git LFS Details
|
|
Git LFS Details
|
docs/figures/pretraining-overview.svg
CHANGED
|
|
|
|
inference_native/__init__.py
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
"""Native inference components for Solstice-1."""
|
|
|
|
| 1 |
+
"""Native inference components for Solstice-1-175M."""
|
provenance/LINEAGE.json
CHANGED
|
@@ -4,7 +4,7 @@
|
|
| 4 |
"OpenSML-150M",
|
| 5 |
"OpenSML-JSONLite-175M"
|
| 6 |
],
|
| 7 |
-
"model": "OpenSML/Solstice-1",
|
| 8 |
"parameter_label": "175M",
|
| 9 |
"parameters": 174202432,
|
| 10 |
"pretraining": {
|
|
|
|
| 4 |
"OpenSML-150M",
|
| 5 |
"OpenSML-JSONLite-175M"
|
| 6 |
],
|
| 7 |
+
"model": "OpenSML/Solstice-1-175M",
|
| 8 |
"parameter_label": "175M",
|
| 9 |
"parameters": 174202432,
|
| 10 |
"pretraining": {
|
scripts/plot_results.py
CHANGED
|
@@ -11,7 +11,7 @@ import numpy as np
|
|
| 11 |
|
| 12 |
ROOT = Path(__file__).resolve().parents[1]
|
| 13 |
OUT = ROOT / 'docs/figures'
|
| 14 |
-
COLORS = {'Solstice-1': '#176b65', 'OpenSML-150M': '#67768a',
|
| 15 |
'SmolLM2-135M Instruct': '#c58a37', 'Checkpoint 167': '#9476a7'}
|
| 16 |
|
| 17 |
|
|
@@ -36,12 +36,12 @@ def main():
|
|
| 36 |
smol = public['HuggingFaceTB/SmolLM2-135M-Instruct']
|
| 37 |
fig, axes = plt.subplots(2, 2, figsize=(11.5, 7.6))
|
| 38 |
fig.subplots_adjust(left=.2, right=.97, top=.84, bottom=.16, wspace=.85, hspace=.55)
|
| 39 |
-
fig.suptitle('Solstice-1 | Knowledge and commonsense', x=.08, ha='left', y=.97,
|
| 40 |
fontsize=19, fontweight='bold')
|
| 41 |
fig.text(.08, .91, 'Zero-shot normalized accuracy 路 higher is better', fontsize=12)
|
| 42 |
for ax, (key, label) in zip(axes.flat, [('arc_easy', 'ARC Easy'),
|
| 43 |
('arc_challenge', 'ARC Challenge'), ('piqa', 'PIQA'), ('hellaswag', 'HellaSwag')]):
|
| 44 |
-
names = ['Solstice-1', 'OpenSML-150M']
|
| 45 |
values = [choice[key]['acc_norm'], old['choices'][key]['acc_norm']]
|
| 46 |
if key in smol['choices']:
|
| 47 |
names.append('SmolLM2-135M Instruct')
|
|
@@ -65,12 +65,12 @@ def main():
|
|
| 65 |
keys = [('strict', 'prompt'), ('strict', 'instruction'), ('loose', 'prompt'), ('loose', 'instruction')]
|
| 66 |
fig, ax = plt.subplots(figsize=(11.5, 5.9))
|
| 67 |
fig.subplots_adjust(left=.08, right=.98, top=.8, bottom=.24)
|
| 68 |
-
fig.suptitle('Solstice-1 | Instruction following', x=.08, ha='left', y=.97,
|
| 69 |
fontsize=19, fontweight='bold')
|
| 70 |
fig.text(.08, .9, 'IFEval accuracy 路 higher is better', fontsize=12)
|
| 71 |
x = np.arange(4)
|
| 72 |
for i, (name, data) in enumerate([('OpenSML-150M', old['ifeval']),
|
| 73 |
-
('Checkpoint 167', prior), ('Solstice-1', own)]):
|
| 74 |
values = [data[k][u+'_accuracy'] for k, u in keys]
|
| 75 |
bars = ax.bar(x+(i-1)*.25, values, width=.22, label=name, color=COLORS[name])
|
| 76 |
ax.bar_label(bars, labels=[f'{v:.2%}' for v in values], padding=4, fontsize=9)
|
|
@@ -90,10 +90,10 @@ def main():
|
|
| 90 |
values = [chat['categories'][key]['answer_match']/chat['categories'][key]['count'] for key, _ in categories]
|
| 91 |
fig, ax = plt.subplots(figsize=(11.5, 5.8))
|
| 92 |
fig.subplots_adjust(left=.25, right=.96, top=.79, bottom=.2)
|
| 93 |
-
fig.suptitle('Solstice-1 | Short-chat diagnostics', x=.08, ha='left', y=.97,
|
| 94 |
fontsize=19, fontweight='bold')
|
| 95 |
fig.text(.08, .9, 'Conservative automatic answer matching 路 authored panel', fontsize=12)
|
| 96 |
-
ax.barh(range(6), values, color=COLORS['Solstice-1'], height=.6)
|
| 97 |
ax.set_yticks(range(6), [label for _, label in categories]); ax.invert_yaxis()
|
| 98 |
ax.set_xlim(0, 1); ax.xaxis.set_major_formatter(PercentFormatter(1, decimals=0))
|
| 99 |
ax.tick_params(axis='y', length=0); ax.grid(axis='x', alpha=.15); ax.set_axisbelow(True)
|
|
|
|
| 11 |
|
| 12 |
ROOT = Path(__file__).resolve().parents[1]
|
| 13 |
OUT = ROOT / 'docs/figures'
|
| 14 |
+
COLORS = {'Solstice-1-175M': '#176b65', 'OpenSML-150M': '#67768a',
|
| 15 |
'SmolLM2-135M Instruct': '#c58a37', 'Checkpoint 167': '#9476a7'}
|
| 16 |
|
| 17 |
|
|
|
|
| 36 |
smol = public['HuggingFaceTB/SmolLM2-135M-Instruct']
|
| 37 |
fig, axes = plt.subplots(2, 2, figsize=(11.5, 7.6))
|
| 38 |
fig.subplots_adjust(left=.2, right=.97, top=.84, bottom=.16, wspace=.85, hspace=.55)
|
| 39 |
+
fig.suptitle('Solstice-1-175M | Knowledge and commonsense', x=.08, ha='left', y=.97,
|
| 40 |
fontsize=19, fontweight='bold')
|
| 41 |
fig.text(.08, .91, 'Zero-shot normalized accuracy 路 higher is better', fontsize=12)
|
| 42 |
for ax, (key, label) in zip(axes.flat, [('arc_easy', 'ARC Easy'),
|
| 43 |
('arc_challenge', 'ARC Challenge'), ('piqa', 'PIQA'), ('hellaswag', 'HellaSwag')]):
|
| 44 |
+
names = ['Solstice-1-175M', 'OpenSML-150M']
|
| 45 |
values = [choice[key]['acc_norm'], old['choices'][key]['acc_norm']]
|
| 46 |
if key in smol['choices']:
|
| 47 |
names.append('SmolLM2-135M Instruct')
|
|
|
|
| 65 |
keys = [('strict', 'prompt'), ('strict', 'instruction'), ('loose', 'prompt'), ('loose', 'instruction')]
|
| 66 |
fig, ax = plt.subplots(figsize=(11.5, 5.9))
|
| 67 |
fig.subplots_adjust(left=.08, right=.98, top=.8, bottom=.24)
|
| 68 |
+
fig.suptitle('Solstice-1-175M | Instruction following', x=.08, ha='left', y=.97,
|
| 69 |
fontsize=19, fontweight='bold')
|
| 70 |
fig.text(.08, .9, 'IFEval accuracy 路 higher is better', fontsize=12)
|
| 71 |
x = np.arange(4)
|
| 72 |
for i, (name, data) in enumerate([('OpenSML-150M', old['ifeval']),
|
| 73 |
+
('Checkpoint 167', prior), ('Solstice-1-175M', own)]):
|
| 74 |
values = [data[k][u+'_accuracy'] for k, u in keys]
|
| 75 |
bars = ax.bar(x+(i-1)*.25, values, width=.22, label=name, color=COLORS[name])
|
| 76 |
ax.bar_label(bars, labels=[f'{v:.2%}' for v in values], padding=4, fontsize=9)
|
|
|
|
| 90 |
values = [chat['categories'][key]['answer_match']/chat['categories'][key]['count'] for key, _ in categories]
|
| 91 |
fig, ax = plt.subplots(figsize=(11.5, 5.8))
|
| 92 |
fig.subplots_adjust(left=.25, right=.96, top=.79, bottom=.2)
|
| 93 |
+
fig.suptitle('Solstice-1-175M | Short-chat diagnostics', x=.08, ha='left', y=.97,
|
| 94 |
fontsize=19, fontweight='bold')
|
| 95 |
fig.text(.08, .9, 'Conservative automatic answer matching 路 authored panel', fontsize=12)
|
| 96 |
+
ax.barh(range(6), values, color=COLORS['Solstice-1-175M'], height=.6)
|
| 97 |
ax.set_yticks(range(6), [label for _, label in categories]); ax.invert_yaxis()
|
| 98 |
ax.set_xlim(0, 1); ax.xaxis.set_major_formatter(PercentFormatter(1, decimals=0))
|
| 99 |
ax.tick_params(axis='y', length=0); ax.grid(axis='x', alpha=.15); ax.set_axisbelow(True)
|
scripts/plot_training.py
CHANGED
|
@@ -36,7 +36,7 @@ def main():
|
|
| 36 |
panel.set_xlim(lower,upper); panel.grid(alpha=.16)
|
| 37 |
panel.axvline(36.801216512,color='#5f6984',lw=.8,ls='--')
|
| 38 |
panel.axvspan(40.001339392,42.001367040,color='#ead9ac',alpha=.4)
|
| 39 |
-
fig.suptitle('Solstice-1
|
| 40 |
fig.supxlabel('Dashed line: Python/SQL introduced 路 Shaded region: final cooldown\nTraining and validation use different data; early training loss above 5.6 is outside the overview scale.' if lower==0 else 'Dashed line: Python/SQL introduced 路 Shaded region: final cooldown\nMetrics are transcribed from saved logs, with their original reported precision.',fontsize=9,color='#44505c')
|
| 41 |
for ext in ('png','svg'): fig.savefig(output/f'{name}.{ext}',dpi=180)
|
| 42 |
plt.close(fig)
|
|
|
|
| 36 |
panel.set_xlim(lower,upper); panel.grid(alpha=.16)
|
| 37 |
panel.axvline(36.801216512,color='#5f6984',lw=.8,ls='--')
|
| 38 |
panel.axvspan(40.001339392,42.001367040,color='#ead9ac',alpha=.4)
|
| 39 |
+
fig.suptitle('Solstice-1-175M | '+('Pretraining history' if lower==0 else 'Late training and 2B-token cooldown'),fontsize=16,fontweight='bold')
|
| 40 |
fig.supxlabel('Dashed line: Python/SQL introduced 路 Shaded region: final cooldown\nTraining and validation use different data; early training loss above 5.6 is outside the overview scale.' if lower==0 else 'Dashed line: Python/SQL introduced 路 Shaded region: final cooldown\nMetrics are transcribed from saved logs, with their original reported precision.',fontsize=9,color='#44505c')
|
| 41 |
for ext in ('png','svg'): fig.savefig(output/f'{name}.{ext}',dpi=180)
|
| 42 |
plt.close(fig)
|
scripts/verify_package.py
CHANGED
|
@@ -37,7 +37,7 @@ def main():
|
|
| 37 |
else:
|
| 38 |
checked += 1
|
| 39 |
elif not args.source_only:
|
| 40 |
-
failures.append('Missing model.safetensors; download it from OpenSML/Solstice-1.')
|
| 41 |
print(json.dumps({'passed': not failures, 'files_verified': checked, 'weights_present': weights.is_file(), 'failures': failures}, indent=2))
|
| 42 |
raise SystemExit(bool(failures))
|
| 43 |
|
|
|
|
| 37 |
else:
|
| 38 |
checked += 1
|
| 39 |
elif not args.source_only:
|
| 40 |
+
failures.append('Missing model.safetensors; download it from OpenSML/Solstice-1-175M.')
|
| 41 |
print(json.dumps({'passed': not failures, 'files_verified': checked, 'weights_present': weights.is_file(), 'failures': failures}, indent=2))
|
| 42 |
raise SystemExit(bool(failures))
|
| 43 |
|