theredteamtech commited on
Commit
bdd65ed
·
verified ·
1 Parent(s): acd5bc2

Add complete JengaCoder v1.2 model card

Browse files
Files changed (1) hide show
  1. README.md +208 -0
README.md ADDED
@@ -0,0 +1,208 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
4
+ pipeline_tag: text-generation
5
+ language:
6
+ - en
7
+ tags:
8
+ - gguf
9
+ - llama-cpp
10
+ - qlora
11
+ - code
12
+ - arduino
13
+ - esp32
14
+ ---
15
+
16
+ # JengaCoder v1.2
17
+
18
+ ---
19
+
20
+ # JengaCoder v1.2
21
+
22
+ JengaCoder v1.2 is a compact offline coding and engineering assistant developed for JengaAI Edge.
23
+
24
+ It is intended for students, teachers and developers using affordable laptops or working in environments where internet connectivity, cloud access and data cost are significant constraints.
25
+
26
+ The model runs locally through `llama.cpp` and does not require a dedicated GPU for inference.
27
+
28
+ ## Model details
29
+
30
+ * **Developer:** Abdulkarim Mohammed Ngulangwa / JengaAI Edge
31
+ * **Base model:** `Qwen/Qwen2.5-Coder-1.5B-Instruct`
32
+ * **Base-model revision:** `2e1fd397ee46e1388853d2af2c993145b0f1098a`
33
+ * **Fine-tuning method:** QLoRA supervised fine-tuning
34
+ * **Parameters:** approximately 1.54 billion
35
+ * **Published format:** GGUF
36
+ * **Quantization:** Q4_K_M
37
+ * **Runtime:** llama.cpp
38
+ * **Primary language:** English
39
+ * **Model file:** `JengaCoder-v1.2-Q4_K_M.gguf`
40
+ * **Model SHA-256:** `e253182086f2bfd9e48ee3e4b683f151276fa1dfc6510faea8be0824dbe433bc`
41
+
42
+ ## Intended uses
43
+
44
+ JengaCoder is intended to provide offline assistance with:
45
+
46
+ * Python and C++ programming
47
+ * debugging and code explanation
48
+ * memory-conscious programming
49
+ * introductory web development
50
+ * Arduino and ESP32 projects
51
+ * embedded-system explanations
52
+ * practical coding and engineering education
53
+
54
+ The model is particularly intended for educational and development environments where continuous cloud access is not practical.
55
+
56
+ ## Running with llama.cpp
57
+
58
+ Download the GGUF file and run:
59
+
60
+ ```bash
61
+ llama-cli \
62
+ -m JengaCoder-v1.2-Q4_K_M.gguf \
63
+ -cnv \
64
+ -n 512 \
65
+ --temp 0.2
66
+ ```
67
+
68
+ An example prompt is:
69
+
70
+ ```text
71
+ Write a complete Python function named count_error_lines that reads a text file too large to fit in memory and returns the number of lines containing the exact substring ERROR. Use only the Python standard library, process the file in one pass with O(1) additional memory, handle FileNotFoundError gracefully, and explain the time and space complexity. Do not store matching lines in a list.
72
+ ```
73
+
74
+ ## Training
75
+
76
+ JengaCoder v1.2 was trained using QLoRA with the following configuration:
77
+
78
+ * training records: 160
79
+ * validation records: 20
80
+ * epochs: 2
81
+ * final global step: 40
82
+ * random seed: 42
83
+ * LoRA rank: 16
84
+ * LoRA alpha: 32
85
+ * LoRA dropout: 0.05
86
+ * training quantization: NF4 with double quantization
87
+ * learning rate: `1e-4`
88
+ * maximum sequence length: 2048
89
+
90
+ The LoRA adapter targeted:
91
+
92
+ * `q_proj`
93
+ * `k_proj`
94
+ * `v_proj`
95
+ * `o_proj`
96
+ * `gate_proj`
97
+ * `up_proj`
98
+ * `down_proj`
99
+
100
+ The authentic final trainer state records:
101
+
102
+ * final evaluation loss: `0.9508873224258423`
103
+ * final evaluation mean token accuracy: `0.8057295680046082`
104
+ * completed epoch: `2.0`
105
+ * completed global step: `40`
106
+
107
+ The original adapter, trainer state, training script, dataset checksums and reproduction documentation are available in the Gate 2 repository.
108
+
109
+ ## Evaluation
110
+
111
+ ### ADTC participant profiler
112
+
113
+ A final participant-mode profiler run on an Intel Core i7-1065G7 laptop with 7.6 GB RAM and no discrete GPU produced:
114
+
115
+ * generation throughput: approximately 19.02 tokens per second
116
+ * peak resident memory: approximately 1692.38 MB
117
+ * ARC Easy normalized accuracy: 0.70
118
+ * ARC Easy samples: 50
119
+ * runtime: CPU-only llama.cpp
120
+
121
+ Performance varies with CPU frequency, background processes, thermal conditions and llama.cpp build configuration.
122
+
123
+ ### Constraint-focused evaluation
124
+
125
+ The repository contains a frozen 10-prompt ConstraintBench diagnostic comparing the Qwen base model and JengaCoder v1.2.
126
+
127
+ | Evaluation | Qwen base | JengaCoder v1.2 |
128
+ | ----------------------------- | --------: | --------------: |
129
+ | Overall ConstraintBench score | 77/100 | 59/100 |
130
+ | Physical/electrical subset | 24/40 | 15/40 |
131
+
132
+ These results are disclosed because they show that fine-tuning did not produce universal improvement. JengaCoder v1.2 showed useful behavior on several practical coding tasks but regressed on some tightly constrained physical-computing questions.
133
+
134
+ Complete prompts, rubrics, outputs and scoring evidence are available in the project repository.
135
+
136
+ ## Limitations
137
+
138
+ JengaCoder is a compact 1.54B-parameter model and has substantially less reasoning capacity than larger language models.
139
+
140
+ Known limitations include:
141
+
142
+ * occasional failure to follow strict hardware constraints
143
+ * possible invention of unrequested components
144
+ * incorrect electrical or sensor assumptions
145
+ * inconsistent handling of impossible requirements
146
+ * possible coding or memory-complexity errors
147
+ * English is the primary supported language
148
+ * generated code may require correction before deployment
149
+
150
+ A known example is that the model may incorrectly recommend driving a bare external LED without a current-limiting resistor. Such advice is unsafe and must not be followed.
151
+
152
+ All hardware guidance should be reviewed against component datasheets and tested safely before deployment.
153
+
154
+ ## Safety recommendations
155
+
156
+ Users should:
157
+
158
+ * verify electrical voltage and current limits
159
+ * use appropriate current-limiting and protection components
160
+ * review generated wiring before powering a circuit
161
+ * test code incrementally
162
+ * avoid using generated output in safety-critical systems without qualified review
163
+ * treat the model as an educational assistant rather than an authoritative engineering source
164
+
165
+ ## Reproducibility and provenance
166
+
167
+ The complete Gate 2 evidence package is available at:
168
+
169
+ https://github.com/theredteamtech/jengaai-edge/tree/gate2
170
+
171
+ It includes:
172
+
173
+ * original LoRA adapter weights
174
+ * adapter configuration
175
+ * final trainer state
176
+ * exact training script
177
+ * training and validation datasets
178
+ * SHA-256 checksum manifest
179
+ * base-model identity
180
+ * merge script
181
+ * quantization script
182
+ * benchmark evidence
183
+ * complete base-versus-fine-tuned output examples
184
+
185
+ ## Dataset and licensing
186
+
187
+ The JengaAI Edge training and validation datasets are included in the GitHub repository:
188
+
189
+ * `training/data/jengacoder_train_v1.json`
190
+ * `training/data/jengacoder_validation_v1.json`
191
+
192
+ The repository code and included datasets are distributed under the repository’s GNU General Public License v3.0.
193
+
194
+ The model is derived from `Qwen/Qwen2.5-Coder-1.5B-Instruct`, which is distributed under the Apache License 2.0. Users must comply with all applicable upstream and repository licensing terms.
195
+
196
+ ## Development status
197
+
198
+ JengaCoder v1.2 is the stable submitted model.
199
+
200
+ Later experimental adapters were not published as replacements because evaluation revealed regressions. A separate corrective v2 dataset is being developed to improve physical grounding, constraint following and embedded-system safety.
201
+
202
+ The v2 development data was not used to train JengaCoder v1.2.
203
+
204
+ ## Contact
205
+
206
+ * **Developer:** Abdulkarim Mohammed Ngulangwa
207
+ * **GitHub:** https://github.com/theredteamtech
208
+ * **Project repository:** https://github.com/theredteamtech/jengaai-edge