wesleysimplicio commited on
Commit
3c725e4
·
verified ·
1 Parent(s): 8b649b0

Sync from GitHub: allowlist, adapter in lora/, card metadata

Browse files

https://github.com/simpletibr/simplicio-27b/commit/e42e364095480d39a2066acd750e28f0371b1b0d

Files changed (4) hide show
  1. LICENSE +202 -0
  2. Modelfile +20 -24
  3. README.md +134 -81
  4. generation_config.json +6 -3
LICENSE ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ Apache License
3
+ Version 2.0, January 2004
4
+ http://www.apache.org/licenses/
5
+
6
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
7
+
8
+ 1. Definitions.
9
+
10
+ "License" shall mean the terms and conditions for use, reproduction,
11
+ and distribution as defined by Sections 1 through 9 of this document.
12
+
13
+ "Licensor" shall mean the copyright owner or entity authorized by
14
+ the copyright owner that is granting the License.
15
+
16
+ "Legal Entity" shall mean the union of the acting entity and all
17
+ other entities that control, are controlled by, or are under common
18
+ control with that entity. For the purposes of this definition,
19
+ "control" means (i) the power, direct or indirect, to cause the
20
+ direction or management of such entity, whether by contract or
21
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
22
+ outstanding shares, or (iii) beneficial ownership of such entity.
23
+
24
+ "You" (or "Your") shall mean an individual or Legal Entity
25
+ exercising permissions granted by this License.
26
+
27
+ "Source" form shall mean the preferred form for making modifications,
28
+ including but not limited to software source code, documentation
29
+ source, and configuration files.
30
+
31
+ "Object" form shall mean any form resulting from mechanical
32
+ transformation or translation of a Source form, including but
33
+ not limited to compiled object code, generated documentation,
34
+ and conversions to other media types.
35
+
36
+ "Work" shall mean the work of authorship, whether in Source or
37
+ Object form, made available under the License, as indicated by a
38
+ copyright notice that is included in or attached to the work
39
+ (an example is provided in the Appendix below).
40
+
41
+ "Derivative Works" shall mean any work, whether in Source or Object
42
+ form, that is based on (or derived from) the Work and for which the
43
+ editorial revisions, annotations, elaborations, or other modifications
44
+ represent, as a whole, an original work of authorship. For the purposes
45
+ of this License, Derivative Works shall not include works that remain
46
+ separable from, or merely link (or bind by name) to the interfaces of,
47
+ the Work and Derivative Works thereof.
48
+
49
+ "Contribution" shall mean any work of authorship, including
50
+ the original version of the Work and any modifications or additions
51
+ to that Work or Derivative Works thereof, that is intentionally
52
+ submitted to Licensor for inclusion in the Work by the copyright owner
53
+ or by an individual or Legal Entity authorized to submit on behalf of
54
+ the copyright owner. For the purposes of this definition, "submitted"
55
+ means any form of electronic, verbal, or written communication sent
56
+ to the Licensor or its representatives, including but not limited to
57
+ communication on electronic mailing lists, source code control systems,
58
+ and issue tracking systems that are managed by, or on behalf of, the
59
+ Licensor for the purpose of discussing and improving the Work, but
60
+ excluding communication that is conspicuously marked or otherwise
61
+ designated in writing by the copyright owner as "Not a Contribution."
62
+
63
+ "Contributor" shall mean Licensor and any individual or Legal Entity
64
+ on behalf of whom a Contribution has been received by Licensor and
65
+ subsequently incorporated within the Work.
66
+
67
+ 2. Grant of Copyright License. Subject to the terms and conditions of
68
+ this License, each Contributor hereby grants to You a perpetual,
69
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
70
+ copyright license to reproduce, prepare Derivative Works of,
71
+ publicly display, publicly perform, sublicense, and distribute the
72
+ Work and such Derivative Works in Source or Object form.
73
+
74
+ 3. Grant of Patent License. Subject to the terms and conditions of
75
+ this License, each Contributor hereby grants to You a perpetual,
76
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
77
+ (except as stated in this section) patent license to make, have made,
78
+ use, offer to sell, sell, import, and otherwise transfer the Work,
79
+ where such license applies only to those patent claims licensable
80
+ by such Contributor that are necessarily infringed by their
81
+ Contribution(s) alone or by combination of their Contribution(s)
82
+ with the Work to which such Contribution(s) was submitted. If You
83
+ institute patent litigation against any entity (including a
84
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
85
+ or a Contribution incorporated within the Work constitutes direct
86
+ or contributory patent infringement, then any patent licenses
87
+ granted to You under this License for that Work shall terminate
88
+ as of the date such litigation is filed.
89
+
90
+ 4. Redistribution. You may reproduce and distribute copies of the
91
+ Work or Derivative Works thereof in any medium, with or without
92
+ modifications, and in Source or Object form, provided that You
93
+ meet the following conditions:
94
+
95
+ (a) You must give any other recipients of the Work or
96
+ Derivative Works a copy of this License; and
97
+
98
+ (b) You must cause any modified files to carry prominent notices
99
+ stating that You changed the files; and
100
+
101
+ (c) You must retain, in the Source form of any Derivative Works
102
+ that You distribute, all copyright, patent, trademark, and
103
+ attribution notices from the Source form of the Work,
104
+ excluding those notices that do not pertain to any part of
105
+ the Derivative Works; and
106
+
107
+ (d) If the Work includes a "NOTICE" text file as part of its
108
+ distribution, then any Derivative Works that You distribute must
109
+ include a readable copy of the attribution notices contained
110
+ within such NOTICE file, excluding those notices that do not
111
+ pertain to any part of the Derivative Works, in at least one
112
+ of the following places: within a NOTICE text file distributed
113
+ as part of the Derivative Works; within the Source form or
114
+ documentation, if provided along with the Derivative Works; or,
115
+ within a display generated by the Derivative Works, if and
116
+ wherever such third-party notices normally appear. The contents
117
+ of the NOTICE file are for informational purposes only and
118
+ do not modify the License. You may add Your own attribution
119
+ notices within Derivative Works that You distribute, alongside
120
+ or as an addendum to the NOTICE text from the Work, provided
121
+ that such additional attribution notices cannot be construed
122
+ as modifying the License.
123
+
124
+ You may add Your own copyright statement to Your modifications and
125
+ may provide additional or different license terms and conditions
126
+ for use, reproduction, or distribution of Your modifications, or
127
+ for any such Derivative Works as a whole, provided Your use,
128
+ reproduction, and distribution of the Work otherwise complies with
129
+ the conditions stated in this License.
130
+
131
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
132
+ any Contribution intentionally submitted for inclusion in the Work
133
+ by You to the Licensor shall be under the terms and conditions of
134
+ this License, without any additional terms or conditions.
135
+ Notwithstanding the above, nothing herein shall supersede or modify
136
+ the terms of any separate license agreement you may have executed
137
+ with Licensor regarding such Contributions.
138
+
139
+ 6. Trademarks. This License does not grant permission to use the trade
140
+ names, trademarks, service marks, or product names of the Licensor,
141
+ except as required for reasonable and customary use in describing the
142
+ origin of the Work and reproducing the content of the NOTICE file.
143
+
144
+ 7. Disclaimer of Warranty. Unless required by applicable law or
145
+ agreed to in writing, Licensor provides the Work (and each
146
+ Contributor provides its Contributions) on an "AS IS" BASIS,
147
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
148
+ implied, including, without limitation, any warranties or conditions
149
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
150
+ PARTICULAR PURPOSE. You are solely responsible for determining the
151
+ appropriateness of using or redistributing the Work and assume any
152
+ risks associated with Your exercise of permissions under this License.
153
+
154
+ 8. Limitation of Liability. In no event and under no legal theory,
155
+ whether in tort (including negligence), contract, or otherwise,
156
+ unless required by applicable law (such as deliberate and grossly
157
+ negligent acts) or agreed to in writing, shall any Contributor be
158
+ liable to You for damages, including any direct, indirect, special,
159
+ incidental, or consequential damages of any character arising as a
160
+ result of this License or out of the use or inability to use the
161
+ Work (including but not limited to damages for loss of goodwill,
162
+ work stoppage, computer failure or malfunction, or any and all
163
+ other commercial damages or losses), even if such Contributor
164
+ has been advised of the possibility of such damages.
165
+
166
+ 9. Accepting Warranty or Additional Liability. While redistributing
167
+ the Work or Derivative Works thereof, You may choose to offer,
168
+ and charge a fee for, acceptance of support, warranty, indemnity,
169
+ or other liability obligations and/or rights consistent with this
170
+ License. However, in accepting such obligations, You may act only
171
+ on Your own behalf and on Your sole responsibility, not on behalf
172
+ of any other Contributor, and only if You agree to indemnify,
173
+ defend, and hold each Contributor harmless for any liability
174
+ incurred by, or claims asserted against, such Contributor by reason
175
+ of your accepting any such warranty or additional liability.
176
+
177
+ END OF TERMS AND CONDITIONS
178
+
179
+ APPENDIX: How to apply the Apache License to your work.
180
+
181
+ To apply the Apache License to your work, attach the following
182
+ boilerplate notice, with the fields enclosed by brackets "[]"
183
+ replaced with your own identifying information. (Don't include
184
+ the brackets!) The text should be enclosed in the appropriate
185
+ comment syntax for the file format. We also recommend that a
186
+ file or class name and description of purpose be included on the
187
+ same "printed page" as the copyright notice for easier
188
+ identification within third-party archives.
189
+
190
+ Copyright [yyyy] [name of copyright owner]
191
+
192
+ Licensed under the Apache License, Version 2.0 (the "License");
193
+ you may not use this file except in compliance with the License.
194
+ You may obtain a copy of the License at
195
+
196
+ http://www.apache.org/licenses/LICENSE-2.0
197
+
198
+ Unless required by applicable law or agreed to in writing, software
199
+ distributed under the License is distributed on an "AS IS" BASIS,
200
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
201
+ See the License for the specific language governing permissions and
202
+ limitations under the License.
Modelfile CHANGED
@@ -1,29 +1,25 @@
1
- # Ollama Modelfile for Simplicio 27B
2
- # Architecture: 27B Dense/Hybrid Foundation Model
3
- # Optimized for: Atomic Search/Replace Diffs & Zero-Token-Waste Execution
4
- # Quantization recommended: Q4_K_M (~16.8 GB) or Q8_0 (~28.5 GB)
5
 
6
- FROM ./simplicio-27b-Q4_K_M.gguf
 
7
 
8
- # Generation Parameters
9
- PARAMETER temperature 0.2
10
- PARAMETER top_p 0.95
11
- PARAMETER top_k 40
12
- PARAMETER repeat_penalty 1.1
13
- PARAMETER num_ctx 32768
14
 
15
- # Deterministic Stop Tokens
16
- PARAMETER stop "<|im_end|>"
17
- PARAMETER stop "</deliver>"
18
- PARAMETER stop "<|endoftext|>"
19
 
20
- # ChatML Template
21
- TEMPLATE """{{ if .System }}<|im_start|>system
22
- {{ .System }}<|im_end|>
23
- {{ end }}{{ if .Prompt }}<|im_start|>user
24
- {{ .Prompt }}<|im_end|>
25
- {{ end }}<|im_start|>assistant
26
- {{ .Response }}<|im_end|>"""
 
27
 
28
- # Official System Prompt
29
- SYSTEM """Voce e o Simplicio 27B, um modelo especializado em engenharia de software autonoma e edicao cirurgica atomica de codigo (SEARCH/REPLACE), treinado pela simpleti.com.br para maxima assertividade e desperdicio zero de tokens."""
 
1
+ # Simplicio 27B for Ollama (the only Modelfile): ollama run wesleysimplicio/simplicio-27b --think=false
2
+ # Rebuild: put Qwen3.8-27B.Q4_K_M.gguf from huggingface.co/wesleysimplicio/Simplicio-27B next to this file.
 
 
3
 
4
+ # One FROM only. The vision projector (Qwen3.8-27B.BF16-mmproj.gguf) is left out on purpose: Simplicio is used for text only.
5
+ FROM ./Qwen3.8-27B.Q4_K_M.gguf
6
 
7
+ # Ollama's built-in Qwen 3.5 prompt renderer and tool-call/thinking parser (Ollama >= 0.30.0).
8
+ RENDERER qwen3.5
9
+ PARSER qwen3.5
10
+ REQUIRES 0.30.0
 
 
11
 
12
+ # Same context as vLLM: deploy/context.env MAX_MODEL_LEN.
13
+ PARAMETER num_ctx 40960
 
 
14
 
15
+ # Qwen3.8-27B recommended sampling for non-thinking mode (huggingface.co/Qwen/Qwen3.8-27B).
16
+ PARAMETER temperature 0.7
17
+ PARAMETER top_p 0.8
18
+ PARAMETER top_k 20
19
+ PARAMETER min_p 0
20
+ PARAMETER presence_penalty 1.5
21
+ PARAMETER repeat_penalty 1
22
+ PARAMETER stop "<|im_end|>"
23
 
24
+ # System prompt used in training (notebooks/Simplicio_27B_Training_Colab.ipynb, system_prompt).
25
+ SYSTEM """Voce e o Simplicio 27B, treinado para executar tarefas de desenvolvimento seguindo rigorosamente os 50 pontos do Simplicio-Loop: Orientacao, Planejamento, Edicao Cirurgica por Diff, Validacao e Entrega Verificada sem alucinacao."""
README.md CHANGED
@@ -3,7 +3,9 @@ language:
3
  - en
4
  - pt
5
  license: apache-2.0
 
6
  base_model: Qwen/Qwen3.8-27B
 
7
  tags:
8
  - qwen
9
  - unsloth
@@ -24,7 +26,7 @@ homepage: https://simpleti.com.br/simplicio-27b/
24
 
25
  <h1 align="center">Simplicio 27B</h1>
26
 
27
- <p align="center">A Qwen3.8-27B fine-tune that answers code-change requests with SEARCH/REPLACE patches instead of whole files.</p>
28
 
29
  <p align="center">
30
  <a href="https://huggingface.co/wesleysimplicio/Simplicio-27B"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Simplicio--27B-yellow.svg" alt="Hugging Face"></a>
@@ -36,90 +38,81 @@ homepage: https://simpleti.com.br/simplicio-27b/
36
 
37
  ## Overview
38
 
39
- Simplicio 27B is a LoRA fine-tune of [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) by Wesley Simplicio at [SimpleTI](https://simpleti.com.br/simplicio-27b/). It answers a code-change request in five tagged phases (`<orient>`, `<plan>`, `<patch>`, `<validate>`, `<deliver>`). The `<patch>` phase holds SEARCH/REPLACE blocks that touch only the lines that change.
40
 
41
  - **Base model:** Qwen3.8-27B, 27.36B parameters, 64 layers that alternate linear attention (DeltaNet) and full attention in a 3:1 pattern.
42
- - **Files on Hugging Face:** the LoRA adapter (0.64 GB), the merged BF16 checkpoint (18 shards, 55.6 GB), and a GGUF Q4_K_M (16.8 GB) with the base model's vision projector (0.93 GB).
43
- - **Status:** research release. It passes 46.7% of the project's own 120 held-out tasks. It has not been scored on public benchmarks.
44
 
45
- ## Results
46
 
47
- ### Held-out set (120 tasks)
48
 
49
- The full BF16 model ran on Google Colab G4 (NVIDIA RTX PRO 6000 Blackwell, 95 GB) against 120 tasks that were not used in training ([`data/unseen_eval_120.json`](https://github.com/simpletibr/simplicio-27b/blob/main/data/unseen_eval_120.json)). Each task got one attempt at temperature 0, and generation stopped at `</deliver>`. A task passes when its unit test passes after the patch is applied. Aggregate results: [`benchmarks/live_colab_g4_bf16_n120.json`](https://github.com/simpletibr/simplicio-27b/blob/main/benchmarks/live_colab_g4_bf16_n120.json).
 
 
 
 
50
 
51
- | Metric | Result |
52
- |---|---:|
53
- | Unit test passes | **56/120 (46.7%)** |
54
- | SEARCH block found, patch applied | 104/120 (86.7%) |
55
- | Patched file parses (AST) | 104/120 (86.7%) |
56
- | No call to a nonexistent API | 120/120 |
57
- | Output tokens per task, mean / max | 69.9 / 103 |
58
- | Wall time for all 120 tasks | 438 s |
59
 
60
- | Category | Tasks | Unit test passes | Patch applied |
61
- |---|---:|---:|---:|
62
- | Surgical diff and AST precision | 30 | 17 | 20 |
63
- | Edge-case correctness | 30 | 6 | 27 |
64
- | Nonexistent and deprecated API traps | 30 | 27 | 30 |
65
- | Adversarial and out-of-distribution | 30 | 6 | 27 |
66
-
67
- The base model has not been run under this protocol yet, so the gain from fine-tuning is not measured.
68
-
69
- ### Withdrawn numbers
70
 
71
- Earlier versions of this card reported 96.5% accuracy, 116 of 120 tasks passed, 480 tokens per task, a "Top 12" leaderboard and per-token prices. None of these came from running the model on those tasks:
72
 
73
- - 116/120 and its McNemar test come from [`benchmarks/prove_benchmark_120.py`](https://github.com/simpletibr/simplicio-27b/blob/main/benchmarks/prove_benchmark_120.py), which simulates model outputs.
74
- - 480 tokens per task comes from a 3-task smoke test on an A100 ([`benchmarks/empirical_a100_results.json`](https://github.com/simpletibr/simplicio-27b/blob/main/benchmarks/empirical_a100_results.json)).
75
- - 96.5% and the leaderboard rows are hard-coded in [`benchmarks/compare_top10_2026.py`](https://github.com/simpletibr/simplicio-27b/blob/main/benchmarks/compare_top10_2026.py).
76
- - The model is not listed on OpenRouter and has no public per-token price.
77
 
78
- The held-out run above replaces them.
79
 
80
- ### Public coding-agent benchmarks
 
 
 
81
 
82
- Simplicio 27B has not been evaluated on SWE-bench, the Aider benchmark, Terminal-Bench or the Artificial Analysis Coding Agent Index. For scale, these are published Coding Agent Index v1.5 results ([Artificial Analysis](https://artificialanalysis.ai/agents/coding-agents/comparisons/claude-code-vs-codex), retrieved 5 October 2026):
83
 
84
- | Agent | Coding Agent Index | DeepSWE v1.1 | Terminal-Bench 4.0 | SWE-Atlas-QnA | Cost per task |
85
- |---|---:|---:|---:|---:|---:|
86
- | Claude Opus 5.5 (max) | 66 | 68% | 63% | 66% | $13.04 |
87
- | Claude Sonnet 5.5 (max) | 68 | 72% | 66% | 67% | $14.19 |
88
- | GPT-6.1 Sol (xhigh) | 63 | 73% | 55% | 61% | $1.04 |
89
- | Simplicio 27B | not evaluated | – | – | – | – |
90
 
91
- ## Quick start
92
 
93
- ### Ollama
94
 
95
- ```bash
96
- ollama run wesleysimplicio/simplicio-27b
 
 
97
  ```
98
 
99
- The `latest` tag holds the Q4_K_M GGUF and the vision projector. It uses temperature 0.2 and a 32,768-token context, and it stops at `<|im_end|>` and `</deliver>`.
100
-
101
- The installer installs Ollama if it is missing, then runs the model:
102
-
103
- ```bash
104
- curl -fsSL https://raw.githubusercontent.com/simpletibr/simplicio-27b/main/install.sh | bash
105
- ```
106
 
107
  ### vLLM (OpenAI-compatible server with tool calls)
108
 
109
  ```bash
 
110
  git clone https://github.com/simpletibr/simplicio-27b
111
  cd simplicio-27b
112
  ./deploy/serve_vllm.sh wesleysimplicio/Simplicio-27B 8000
113
  ```
114
 
115
- This serves the merged BF16 checkpoint. The weights alone take 55.6 GB; the evaluation above ran on a 95 GB GPU. The script sets:
 
 
 
 
 
 
 
 
116
 
117
- - `--max-model-len 40960`, defined once in [`deploy/context.env`](https://github.com/simpletibr/simplicio-27b/blob/main/deploy/context.env): a measured 31,692-token OpenCode prompt plus 4,096 output tokens.
118
- - `--chat-template` with [`deploy/chat_template_chatml.jinja`](https://github.com/simpletibr/simplicio-27b/blob/main/deploy/chat_template_chatml.jinja). It prefills `<think>` so that `--reasoning-parser qwen3` moves reasoning out of `content`.
119
- - `--enable-auto-tool-choice --tool-call-parser simplicio`, using [`deploy/simplicio_tool_parser.py`](https://github.com/simpletibr/simplicio-27b/blob/main/deploy/simplicio_tool_parser.py). It turns `<tool><name>…</name><params>…</params></tool>` into a single `tool_calls` entry.
120
- - `--served-model-name simplicio-27b simpleti/simplicio-27b`.
121
 
122
- On a smaller GPU, [`Simplicio_27B_Serve_Colab.ipynb`](https://github.com/simpletibr/simplicio-27b/blob/main/Simplicio_27B_Serve_Colab.ipynb) serves the 4-bit base with the LoRA adapter on Colab.
123
 
124
  ### OpenCode
125
 
@@ -143,26 +136,37 @@ Add the vLLM server to `opencode.json` as an OpenAI-compatible provider:
143
  opencode -m simplicio/simplicio-27b
144
  ```
145
 
146
- OpenCode works through tool calls, so point it at the vLLM server. The Ollama template does not declare tools.
147
 
148
- ### Python (Unsloth)
149
 
150
- This loads the adapter on its 4-bit base, the same way [`Simplicio_27B_Merge_Colab.ipynb`](https://github.com/simpletibr/simplicio-27b/blob/main/Simplicio_27B_Merge_Colab.ipynb) does:
151
 
152
  ```python
153
- from unsloth import FastLanguageModel
 
154
 
155
- model, tokenizer = FastLanguageModel.from_pretrained(
156
- model_name="wesleysimplicio/Simplicio-27B", # adapter; the base comes from adapter_config.json
157
- max_seq_length=16384,
158
- load_in_4bit=True,
159
- )
160
- FastLanguageModel.for_inference(model)
161
 
162
  messages = [{"role": "user", "content": "In api/schemas/user.py, accept tax_id with punctuation such as 123.456.789-00."}]
163
- inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to("cuda")
164
- output = model.generate(inputs, max_new_tokens=1024, do_sample=False)
165
- print(tokenizer.decode(output[0][inputs.shape[1]:], skip_special_tokens=True))
 
 
 
 
 
 
 
 
 
 
 
 
 
166
  ```
167
 
168
  ## Output format
@@ -209,12 +213,12 @@ The SEARCH/REPLACE markers are four characters long (`<<<<`, `====`, `>>>>`), no
209
 
210
  ## Training
211
 
212
- The published adapter was produced by [`Simplicio_27B_Training_Colab.ipynb`](https://github.com/simpletibr/simplicio-27b/blob/main/Simplicio_27B_Training_Colab.ipynb). The settings below come from that notebook and the published `adapter_config.json`.
213
 
214
  | Setting | Value |
215
  |---|---|
216
  | Method | QLoRA with Unsloth; base loaded in 4-bit |
217
- | LoRA | r 32, alpha 32, dropout 0, on the q, k, v, o, gate, up and down projections of every layer |
218
  | Steps | 120 steps, batch size 1, gradient accumulation 8 |
219
  | Optimizer | AdamW 8-bit, learning rate 2e-4, cosine schedule, 10 warmup steps, weight decay 0.01 |
220
  | Sequence length | 4,096 |
@@ -222,27 +226,76 @@ The published adapter was produced by [`Simplicio_27B_Training_Colab.ipynb`](htt
222
  | Data | 101 examples in Portuguese: 1 written by hand and 100 generated from a short list of stack and task templates |
223
  | Hardware | Google Colab A100 (40 GB) |
224
 
225
- [`train_simplicio_27b.py`](https://github.com/simpletibr/simplicio-27b/blob/main/train_simplicio_27b.py) is a script version with extra options: freezing the bottom layers, attention-only LoRA, and registering the phase tags as special tokens. The published adapter used none of them, and its tokenizer has no added tokens. [`generate_dataset.py`](https://github.com/simpletibr/simplicio-27b/blob/main/generate_dataset.py) writes `data/simplicio_loop_50pts_train.jsonl` (80 examples) and `data/simplicio_loop_50pts_val.jsonl` (15 examples).
226
 
227
  ## Limitations
228
 
229
- - It passes 46.7% of the held-out tasks, and only 6 of 30 in both the edge-case and the adversarial categories.
230
- - The training set is small (101 examples), templated and in Portuguese. The model follows the format more reliably than it solves the task.
231
  - In the training examples, `<validate>` and `<deliver>` contain written-out results such as "2 passed" or "COMMIT_READY". The model writes these without running anything. Treat them as claims and run your own tests.
232
  - It has not been compared with the base model under the same protocol, and it has not been run on public benchmarks.
 
 
233
  - Aider: the patch markers differ from Aider's edit format, and Aider has not been tested.
234
- - Vision: the GGUF ships the base model's vision projector. Training was text-only, and image input has not been evaluated.
235
 
236
  ## Repository
237
 
238
  | Path | Contents |
239
  |---|---|
240
- | `Simplicio_27B_Training_Colab.ipynb` | Training run that produced the adapter |
241
- | `Simplicio_27B_Merge_Colab.ipynb` | Merges the adapter into 16-bit weights and exports the GGUF Q4_K_M |
242
- | `Simplicio_27B_Serve_Colab.ipynb`, `deploy/` | vLLM serving, chat template, tool parser, context length, Ollama `Modelfile` |
243
- | `data/unseen_eval_120.json` | The 120 held-out tasks |
244
- | `benchmarks/live_colab_g4_bf16_n120.json` | The results above |
245
- | `tests/` | Tests for the serving code: `python -m pytest tests` |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
246
 
247
  ## Citation
248
 
 
3
  - en
4
  - pt
5
  license: apache-2.0
6
+ library_name: transformers
7
  base_model: Qwen/Qwen3.8-27B
8
+ base_model_relation: finetune
9
  tags:
10
  - qwen
11
  - unsloth
 
26
 
27
  <h1 align="center">Simplicio 27B</h1>
28
 
29
+ <p align="center">A Qwen3.8-27B fine-tune trained to answer code-change requests with SEARCH/REPLACE patches instead of whole files.</p>
30
 
31
  <p align="center">
32
  <a href="https://huggingface.co/wesleysimplicio/Simplicio-27B"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Simplicio--27B-yellow.svg" alt="Hugging Face"></a>
 
38
 
39
  ## Overview
40
 
41
+ Simplicio 27B is a LoRA fine-tune of [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) by Wesley Simplicio at [SimpleTI](https://simpleti.com.br/simplicio-27b/). It was trained to answer a code-change request in five tagged phases (`<orient>`, `<plan>`, `<patch>`, `<validate>`, `<deliver>`). The `<patch>` phase holds SEARCH/REPLACE blocks that touch only the lines that change.
42
 
43
  - **Base model:** Qwen3.8-27B, 27.36B parameters, 64 layers that alternate linear attention (DeltaNet) and full attention in a 3:1 pattern.
44
+ - **Files on Hugging Face:** the LoRA adapter in `lora/` (0.64 GB), the merged BF16 checkpoint (18 shards, 55.6 GB), and a GGUF Q4_K_M (16.8 GB) with the base model's vision projector (0.93 GB). Image input is inherited from the base model and has not been evaluated.
45
+ - **Status:** research release. It has not been scored on a reproducible benchmark yet. See [Benchmarks](#benchmarks).
46
 
47
+ ## Benchmarks
48
 
49
+ No benchmark result is published in this README yet. Numbers that were not produced by a reproducible run have been removed.
50
 
51
+ In accordance with the pre-registered protocol (ADR 0001):
52
+ - **Pre-registration:** The evaluation protocol is specified in [`PREREG_eval.md`](PREREG_eval.md) with SHA-256 locked in [`PREREG_sha256.txt`](PREREG_sha256.txt).
53
+ - **Audit template:** Release evaluations are published in [`RELEASE_EVAL.md`](RELEASE_EVAL.md) with two-sided 95% Wilson confidence intervals.
54
+ - **Harness:** The benchmark harness is in [`benchmarks/harness/`](https://github.com/simpletibr/simplicio-27b/tree/main/benchmarks/harness). It runs 40 SEARCH/REPLACE tasks against any OpenAI-compatible endpoint. A task passes only when the hidden pytest tests pass on the patched code. A result will be published here only together with its commit, the model file and its checksum, the prompt hash, the decoding settings, and the per-task output.
55
+ - **Quantization Quality Gate:** Quantized builds (FP8, GGUF, AWQ, MLX) are verified against the unquantized baseline via [`benchmarks/quantized_eval.py`](benchmarks/quantized_eval.py) with Wilson 95% bounds.
56
 
57
+ Not evaluated: SWE-bench, Aider, Terminal-Bench, the Artificial Analysis indices, and image input.
 
 
 
 
 
 
 
58
 
59
+ ## Quick start
 
 
 
 
 
 
 
 
 
60
 
61
+ ### Ollama
62
 
63
+ ```bash
64
+ ollama run wesleysimplicio/simplicio-27b --think=false
65
+ ```
 
66
 
67
+ If Ollama is not installed, get it from [ollama.com/download](https://ollama.com/download). The `latest` tag is built from the root [`Modelfile`](https://github.com/simpletibr/simplicio-27b/blob/main/Modelfile) and needs Ollama 0.30.0 or newer. The Modelfile uses the Q4_K_M GGUF and leaves the base model's vision projector out on purpose: Simplicio is used for text only. The published tag's manifest lists a projector layer too, so the two do not match yet. Image input has not been evaluated. It sets:
68
 
69
+ - `RENDERER qwen3.5` and `PARSER qwen3.5`: Ollama's built-in Qwen 3.5 prompt format. Tool calls come back in `message.tool_calls` and reasoning in `message.thinking`.
70
+ - `num_ctx 40960`: a 40,960-token context, the same as vLLM ([`deploy/context.env`](https://github.com/simpletibr/simplicio-27b/blob/main/deploy/context.env)).
71
+ - Qwen's recommended sampling for Qwen3.8-27B in non-thinking mode: `temperature 0.7`, `top_p 0.8`, `top_k 20`, `min_p 0`, `presence_penalty 1.5`, `repeat_penalty 1`. The benchmark harness uses greedy decoding, so to match it send `"options": {"temperature": 0}`. The tag stops only at `<|im_end|>`; add `"stop": ["</deliver>"]` inside `options` to also stop at the end of the answer.
72
+ - The system prompt used in training (see "Prompt format" below).
73
 
74
+ **Thinking.** The training data has no `<think>` blocks, so run with thinking off. Ollama turns thinking on by default for models that support it, so turn it off on every request:
75
 
76
+ | Client | Thinking off |
77
+ |---|---|
78
+ | `ollama run` | `--think=false`, or `/set nothink` inside a session |
79
+ | `/api/chat` | `"think": false` |
80
+ | `/v1/chat/completions` | `"reasoning_effort": "none"` |
 
81
 
82
+ With thinking on, the model reasons first and the reasoning goes to `message.thinking`. That mode has not been evaluated.
83
 
84
+ **Prompt format.** Training used this system prompt and user message. Ollama adds the system prompt by itself; with vLLM or transformers, send it as the `system` message:
85
 
86
+ ```text
87
+ system: Voce e o Simplicio 27B, treinado para executar tarefas de desenvolvimento seguindo rigorosamente os 50 pontos do Simplicio-Loop: Orientacao, Planejamento, Edicao Cirurgica por Diff, Validacao e Entrega Verificada sem alucinacao.
88
+ user: Contexto: <files, stack and bug report>
89
+ Tarefa: <the change to make>
90
  ```
91
 
92
+ **Not evaluated.** The Q4_K_M file has not been scored. Image input has not been evaluated. Training was text-only. The training data has no tool calls, so tool calling is whatever the base model does.
 
 
 
 
 
 
93
 
94
  ### vLLM (OpenAI-compatible server with tool calls)
95
 
96
  ```bash
97
+ pip install vllm==0.31.0
98
  git clone https://github.com/simpletibr/simplicio-27b
99
  cd simplicio-27b
100
  ./deploy/serve_vllm.sh wesleysimplicio/Simplicio-27B 8000
101
  ```
102
 
103
+ This serves the merged BF16 checkpoint. The weights alone take 55.6 GB. [`deploy/serve_vllm.sh`](https://github.com/simpletibr/simplicio-27b/blob/main/deploy/serve_vllm.sh) is the only serve command, and the Colab notebook runs it too. It sets:
104
+
105
+ - `--max-model-len 40960`, defined once in [`deploy/context.env`](https://github.com/simpletibr/simplicio-27b/blob/main/deploy/context.env): the total context window.
106
+ - No `--chat-template`: vLLM uses the model's own `chat_template.jinja`.
107
+ - `--default-chat-template-kwargs '{"enable_thinking": false}'`: thinking is off by default because the training data has no `<think>` blocks. A request can turn it on with `"chat_template_kwargs": {"enable_thinking": true}`.
108
+ - `--enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3`: vLLM's built-in parsers for Qwen's native `<tool_call><function=…>` format.
109
+ - `--enable-force-include-usage`: every response carries `usage`.
110
+ - `--served-model-name simplicio-27b simpleti/simplicio-27b`: both ids reach the same fine-tuned weights.
111
+ - Listens on `127.0.0.1` by default. To serve other machines, put an HTTPS proxy in front that forwards only `GET /v1/models` and `POST /v1/chat/completions`, and set `VLLM_API_KEY=<random token>` so vLLM requires `Authorization: Bearer <token>` on `/v1/*`.
112
 
113
+ `vllm serve` has no stop flag. `<|im_end|>` already ends generation through `generation_config.json`; to end at `</deliver>`, send `"stop": ["</deliver>"]` in the request.
 
 
 
114
 
115
+ [`notebooks/Simplicio_27B_Serve_Colab.ipynb`](https://github.com/simpletibr/simplicio-27b/blob/main/notebooks/Simplicio_27B_Serve_Colab.ipynb) runs the same script on a Colab G4 (RTX PRO 6000, 96 GB) behind the completion gate and a Cloudflare tunnel.
116
 
117
  ### OpenCode
118
 
 
136
  opencode -m simplicio/simplicio-27b
137
  ```
138
 
139
+ OpenCode works through tool calls, so point it at the vLLM server. OpenCode has not been tested against the Ollama tag.
140
 
141
+ ### Python (Transformers)
142
 
143
+ The Hugging Face repository root is a plain Transformers checkpoint: the merged BF16 weights (55.6 GB). `do_sample=False` is greedy decoding, the setting the benchmark harness uses; without it, `generate` uses the sampling defaults in `generation_config.json`, the same values as the Ollama `Modelfile`.
144
 
145
  ```python
146
+ import torch
147
+ from transformers import AutoModelForImageTextToText, AutoTokenizer
148
 
149
+ repo = "wesleysimplicio/Simplicio-27B"
150
+ tokenizer = AutoTokenizer.from_pretrained(repo)
151
+ model = AutoModelForImageTextToText.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")
 
 
 
152
 
153
  messages = [{"role": "user", "content": "In api/schemas/user.py, accept tax_id with punctuation such as 123.456.789-00."}]
154
+ inputs = tokenizer.apply_chat_template(
155
+ messages, add_generation_prompt=True, enable_thinking=False, return_tensors="pt"
156
+ ).to(model.device)
157
+ output = model.generate(**inputs, max_new_tokens=1024, do_sample=False)
158
+ print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
159
+ ```
160
+
161
+ The LoRA adapter is in [`lora/`](https://huggingface.co/wesleysimplicio/Simplicio-27B/tree/main/lora). To apply it to your own copy of the base model, load the base with `AutoModelForImageTextToText`, the class whose module names (`model.language_model.layers.*`) match the adapter:
162
+
163
+ ```python
164
+ import torch
165
+ from peft import PeftModel
166
+ from transformers import AutoModelForImageTextToText
167
+
168
+ base = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen3.8-27B", dtype=torch.bfloat16, device_map="auto")
169
+ model = PeftModel.from_pretrained(base, "wesleysimplicio/Simplicio-27B", subfolder="lora")
170
  ```
171
 
172
  ## Output format
 
213
 
214
  ## Training
215
 
216
+ The published adapter was produced by [`notebooks/Simplicio_27B_Training_Colab.ipynb`](https://github.com/simpletibr/simplicio-27b/blob/main/notebooks/Simplicio_27B_Training_Colab.ipynb). The settings below come from that notebook and the published `lora/adapter_config.json`.
217
 
218
  | Setting | Value |
219
  |---|---|
220
  | Method | QLoRA with Unsloth; base loaded in 4-bit |
221
+ | LoRA | r 32, alpha 32, dropout 0. q, k, v, o projections on the 16 full-attention layers (3, 7, …, 63); gate, up, down projections on all 64 layers. The `linear_attn` projections of the 48 DeltaNet layers and the vision tower were not adapted (512 LoRA tensors in the published adapter) |
222
  | Steps | 120 steps, batch size 1, gradient accumulation 8 |
223
  | Optimizer | AdamW 8-bit, learning rate 2e-4, cosine schedule, 10 warmup steps, weight decay 0.01 |
224
  | Sequence length | 4,096 |
 
226
  | Data | 101 examples in Portuguese: 1 written by hand and 100 generated from a short list of stack and task templates |
227
  | Hardware | Google Colab A100 (40 GB) |
228
 
229
+ An earlier script, [`train_simplicio_27b.py`](https://github.com/simpletibr/simplicio-27b/blob/b8567df/train_simplicio_27b.py), was removed because its defaults do not reproduce the published adapter: 150 steps, the bottom 48 layers frozen, and the 10 phase tags added as special tokens. The published adapter used none of these, and its tokenizer has no added tokens. [`generate_dataset.py`](https://github.com/simpletibr/simplicio-27b/blob/main/generate_dataset.py) writes `data/simplicio_loop_50pts_train.jsonl` (80 examples) and `data/simplicio_loop_50pts_val.jsonl` (15 examples).
230
 
231
  ## Limitations
232
 
233
+ - No functional correctness of the patches has been measured yet.
234
+ - The training set is small (101 examples), templated and in Portuguese. Whether outputs follow the five-phase format was not measured.
235
  - In the training examples, `<validate>` and `<deliver>` contain written-out results such as "2 passed" or "COMMIT_READY". The model writes these without running anything. Treat them as claims and run your own tests.
236
  - It has not been compared with the base model under the same protocol, and it has not been run on public benchmarks.
237
+ - No checkpoint has been evaluated yet. The BF16 weights, the Q4_K_M GGUF, the Ollama template and parameters, and the 4-bit base + LoRA path are all unmeasured.
238
+ - Decoding: the benchmark harness is greedy (temperature 0). The sampling defaults in the Hugging Face `generation_config.json` and in the Ollama tag are not greedy and have not been measured.
239
  - Aider: the patch markers differ from Aider's edit format, and Aider has not been tested.
240
+ - Vision: the Hugging Face repository ships the base model's vision projector as a separate file. The Modelfile leaves it out, but the published Ollama manifest lists it. Training was text-only, and image input has not been evaluated.
241
 
242
  ## Repository
243
 
244
  | Path | Contents |
245
  |---|---|
246
+ | `scripts/hf_publish.py` | Syncs the Hugging Face repo with this one: file allowlist, adapter in `lora/`, one commit that only adds and updates files, never deletes on the Hub; dry-run by default, `--publish` needs `HF_TOKEN` |
247
+ | `lora/` | Where the LoRA adapter goes (`adapter_config.json`, `adapter_model.safetensors`); the weights are not in git |
248
+ | `notebooks/Simplicio_27B_Training_Colab.ipynb` | Training run that produced the adapter |
249
+ | `notebooks/Simplicio_27B_Merge_Colab.ipynb` | Merges the adapter into 16-bit weights and exports the GGUF Q4_K_M |
250
+ | `notebooks/Simplicio_27B_Serve_Colab.ipynb`, `deploy/` | vLLM serve script, Colab launcher, completion gate, context length |
251
+ | `gateway/` | Gateway code kept in the repo: `set_upstream.php`, `upstream_lease.php` (heartbeat lease: HTTP 503 when the Colab stops renewing), `status.php` (`GET /v1/status`), and `build_catalog.py`, which writes `models.json` (`GET /v1/models`) |
252
+ | `Modelfile` | The Ollama tag `wesleysimplicio/simplicio-27b` |
253
+ | `data/unseen_eval_120.json` | The 120 rows of the held-out set (40 unique tasks) |
254
+ | `benchmarks/harness/` | Benchmark harness: 40 SEARCH/REPLACE tasks with hidden tests |
255
+ | `tests/` | Tests for the serving code: `python3 -m unittest discover -s tests` |
256
+
257
+ ## Development
258
+
259
+ `make check` needs `python3`, `ruff`, `shellcheck`, `php` and `make` on the `PATH`.
260
+
261
+ ```bash
262
+ make check # unittest, ruff F/E9, shellcheck, php -l, vLLM args
263
+ git config core.hooksPath .githooks # once per clone: run make check before every push
264
+ ```
265
+
266
+ The last step parses the exact `vllm serve` argv from `deploy/serve_vllm.sh` with vLLM's own CLI parser, so it needs `vllm==0.31.0` importable. Without vLLM installed it fails with exit code 1 on purpose: a flag that vLLM does not accept must not pass the gate unnoticed. On a machine without vLLM, export `SIMPLICIO_SKIP_VLLM_ARGS=1` to skip only that check (it prints `SKIP: vllm not importable; argv extracted, not validated`); the other four steps still run. There is no CI: all checks run locally. Commit subjects are one plain sentence, without a `type:` prefix. Colab notebooks live in `notebooks/`; `tests/test_layout.py` fails on a notebook at the repository root or on a `blob/main` link to a missing file.
267
+
268
+
269
+ ## Agent integration (MCP)
270
+
271
+ Simplicio 27B provides a native Model Context Protocol (MCP) server over stdio in [`scripts/mcp_server.py`](scripts/mcp_server.py) with zero external dependencies.
272
+
273
+ ### Tools
274
+
275
+ - `apply_surgical_diff`: Atomically applies SEARCH/REPLACE diff blocks to target files with rollback on mismatch.
276
+ - `validate_patch`: Validates diff block structure and syntax offline without modifying files.
277
+ - `format_prompt`: Renders structured prompts conforming to the Simplicio-Loop (`<orient>`, `<plan>`, `<patch>`, `<validate>`, `<deliver>`).
278
+
279
+ ### Diff format compatibility
280
+
281
+ The server natively accepts both standard diff formats:
282
+ - **Aider standard (7-char):** `<<<<<<< SEARCH`, `=======`, `>>>>>>> REPLACE`
283
+ - **Simplicio compact (4-char):** `<<<<`, `====`, `>>>>`
284
+
285
+ ### MCP configuration
286
+
287
+ Add to your MCP client configuration (Claude Desktop, Claude Code, Cursor):
288
+
289
+ ```json
290
+ {
291
+ "mcpServers": {
292
+ "simplicio": {
293
+ "command": "python3",
294
+ "args": ["scripts/mcp_server.py"]
295
+ }
296
+ }
297
+ }
298
+ ```
299
 
300
  ## Citation
301
 
generation_config.json CHANGED
@@ -6,8 +6,11 @@
6
  248044
7
  ],
8
  "pad_token_id": 248044,
9
- "temperature": 1.0,
 
 
10
  "top_k": 20,
11
- "top_p": 0.95,
12
- "transformers_version": "5.5.0"
 
13
  }
 
6
  248044
7
  ],
8
  "pad_token_id": 248044,
9
+ "transformers_version": "5.5.0",
10
+ "temperature": 0.7,
11
+ "top_p": 0.8,
12
  "top_k": 20,
13
+ "min_p": 0.0,
14
+ "presence_penalty": 1.5,
15
+ "repetition_penalty": 1.0
16
  }