faxenoff commited on
Commit
248c7aa
Β·
verified Β·
1 Parent(s): 6ea280f

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +99 -1
README.md CHANGED
@@ -1,3 +1,101 @@
1
  ---
2
- license: mit
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ license: apache-2.0
3
+ language:
4
+ - en
5
+ - ru
6
+ tags:
7
+ - code
8
+ - documentation
9
+ - summarization
10
+ - distillation
11
+ - qwen3
12
+ - gguf
13
+ - llama.cpp
14
+ pipeline_tag: text-generation
15
+ base_model: Qwen/Qwen3-4B
16
+ library_name: gguf
17
  ---
18
+
19
+ # code-daemon-summary-v1
20
+
21
+ Compact **bilingual (English / Russian) code-documentation generator** β€” a 4B GGUF model that
22
+ writes:
23
+
24
+ - **one-sentence entity descriptions** for a source file (functions, classes, fields) as a
25
+ clean markdown bullet list;
26
+ - **module overviews** β€” short prose plus ASCII architecture / data-flow diagrams;
27
+ - **hierarchical codebase summaries** β€” subsystem, product-level and whole-project digests
28
+ built from smaller summaries.
29
+
30
+ Runs anywhere llama.cpp runs; `Q4_K_M` fits in ~2.5 GB. The output language follows the
31
+ request β€” both languages were distilled first-class, not translated.
32
+
33
+ | Task | Output shape |
34
+ |---|---|
35
+ | Entity documentation | `- **Name**: one-sentence description.` per entity |
36
+ | Module overview | `## Overview` prose + `## Architecture` / `## Flow` ASCII diagrams |
37
+ | Hierarchical summaries | paragraph-length subsystem/product/strategy digests |
38
+
39
+ ## Quick start (llama.cpp)
40
+
41
+ The bundled chat template already pins the model's **non-thinking mode** β€” use it as a normal
42
+ ChatML model. If you build raw prompts yourself, end them with the assistant tag followed by an
43
+ empty think block (`<|im_start|>assistant\n<think>\n\n</think>\n\n`) β€” that is the format the
44
+ model was trained with. Greedy decoding (temperature 0) recommended; stop on `<|im_end|>`.
45
+
46
+ ```bash
47
+ llama-cli -m code-daemon-summary-v1-Q4_K_M.gguf -c 8192 --temp 0 \
48
+ -p '<|im_start|>system
49
+ You write one-sentence descriptions for code entities of a single file. Output ONLY a markdown bullet list, ONE bullet per entity: - **<EntityName>**: <one-sentence description>.<|im_end|>
50
+ <|im_start|>user
51
+ Entities: parseArray, encodeValue. File excerpt: <...><|im_end|>
52
+ <|im_start|>assistant
53
+ <think>
54
+
55
+ </think>
56
+
57
+ '
58
+ ```
59
+
60
+ ## Evaluation
61
+
62
+ Held-out prompts via a deterministic content-hash split (provably never trained on), sampled
63
+ across the full bilingual corpus; reference = the 7B teacher's output on the same prompts.
64
+
65
+ | Slice | n | ROUGE-L | sem-cos* | empty outputs |
66
+ |---|---|---|---|---|
67
+ | Entity docs, mixed EN+RU | 300 | 0.618 | **0.904** | 0 |
68
+ | Entity docs, EN-only sample | 300 | 0.636 | 0.897 | 0 |
69
+ | Hierarchical summaries | 7 | 0.331 | 0.893 | 0 |
70
+
71
+ \* all-MiniLM-L6-v2 cosine between student and teacher outputs β€” the paraphrase-aware signal;
72
+ token metrics understate free-form prose.
73
+
74
+ **Language fidelity:** 56/57 Russian-reference examples answered in Russian, 242/243 English in
75
+ English. Residual gap to the teacher: slightly more verbose (len-ratio ~1.3–1.6).
76
+
77
+ ## How it was made
78
+
79
+ - **Base:** [`Qwen/Qwen3-4B`](https://huggingface.co/Qwen/Qwen3-4B) (Apache-2.0), 36 layers,
80
+ ChatML, 151 936-token vocab.
81
+ - **Teacher:** [`Qwen/Qwen2.5-7B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct).
82
+ - **Method:** sequence-level knowledge distillation (SeqKD) β€” LoRA SFT on ~11K teacher
83
+ `(prompt β†’ response)` traces, 3 epochs, prompt tokens masked, then merged into the base.
84
+ An on-policy DistiLLM-2 refinement pass was also trained and **rejected on evaluation**
85
+ (it degraded long-prompt behaviour); v1 is the SeqKD checkpoint.
86
+ - The model powers the long-output documentation stages of a code-intelligence daemon; it is a
87
+ purpose-built component, not a general assistant β€” outside this task distribution its
88
+ behaviour is undefined.
89
+
90
+ ## Training data
91
+
92
+ Teacher traces generated over a mixed corpus: ~90 open-source repositories (Zig, C/C++, C#,
93
+ TypeScript/JS, Python, Go, Rust, Kotlin, Swift, Java) for English and a commercial C#/TS/Python
94
+ codebase for Russian. ~11K kept traces after dedup and corruption filtering. No third-party
95
+ labeled dataset is used.
96
+
97
+ ## License & attribution
98
+
99
+ **Apache-2.0** β€” matches the Qwen3-4B base and the Qwen2.5-7B-Instruct teacher (both Alibaba /
100
+ Qwen team, Apache-2.0). Not legal advice. Base and teacher Β© the Qwen team; please also honour
101
+ their model cards.