radiosilent commited on
Commit
054aeaa
·
verified ·
1 Parent(s): 22a026b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +24 -6
README.md CHANGED
@@ -1,5 +1,8 @@
1
  ---
2
  pipeline_tag: text-generation
 
 
 
3
  tags:
4
  - sysml-v2
5
  - mbestudio
@@ -8,16 +11,31 @@ tags:
8
 
9
  # MBEstudio SysML Assistant
10
 
11
- **Status: preparation only. No trained weights or training data have been uploaded.**
12
 
13
- This private repository is reserved for a future MBEstudio-focused SysML v2 LoRA adapter. The training workflow is being prepared in the MBEstudio source repository and is designed to use a human-reviewed dataset with a held-out evaluation set.
 
 
14
 
15
  ## Intended use
16
 
17
- Assist with drafting, explaining, and repairing SysML v2 models for use with MBEstudio. Generated models must be checked with the MBEstudio compiler and reviewed for engineering correctness.
 
 
 
 
 
 
 
 
 
 
 
 
 
18
 
19
- ## Training and evaluation
20
 
21
- The intended baseline is a compatible Qwen model. Before training, the model owner will confirm the base model license, data rights, hardware requirements, and cost. Evaluation will include compiler acceptance, requirement correctness, repair success, unsupported claims, and target-hardware latency.
22
 
23
- This repository is private. Do not add customer, confidential, or proprietary training data unless it has been explicitly approved for this purpose.
 
1
  ---
2
  pipeline_tag: text-generation
3
+ base_model: Qwen/Qwen3-32B
4
+ language:
5
+ - en
6
  tags:
7
  - sysml-v2
8
  - mbestudio
 
11
 
12
  # MBEstudio SysML Assistant
13
 
14
+ **Status: preparation only. No custom weights have been trained or uploaded.**
15
 
16
+ A planned MBEstudio-focused SysML v2 model. The target base is Qwen3-32B, a 32.8B dense Apache-2.0 model. Qwen3 is multilingual; the MBEstudio fine-tuning corpus is intended to contain English-language instructions and SysML only, with Chinese and other non-English examples excluded.
17
+
18
+ This Hub repository is private while data rights, model quality, and deployment are reviewed. Do not treat this card as a claim that the model has been trained. The current mbes-qwen3-14b-r1 demo served by Studio is a separate, older adapter.
19
 
20
  ## Intended use
21
 
22
+ Assist with drafting, explaining, and repairing SysML v2 models for MBEstudio. Generated models must pass the MBEstudio compiler and engineering review before use.
23
+
24
+ ## Base model and hardware target
25
+
26
+ - Base model: [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B), 32.8B dense parameters, Apache-2.0.
27
+ - Fine-tuning data policy: English-only SysML instructions and outputs; no Chinese or other CJK examples.
28
+ - Local inference target: RTX 5080 with 16 GB VRAM and 64 GB system RAM, using a quantized build with CPU/RAM offload if needed. Latency and context capacity must be measured before release.
29
+ - Training: a GPU training run is required. No paid run has been started.
30
+
31
+ ## Data and validation status
32
+
33
+ The private dataset contains 160 approved training rows and 2 held-out evaluation models. Another 634 public-source rows are candidates only. A compiler screen found 25/103 whole-model candidates and 54/531 declaration candidates with zero diagnostics as standalone files. Many candidate prompts contain Markdown or README instructions; these candidates are not approved for training until the briefs and project-context compiler checks are repaired and reviewed. The proprietary examples and dataset remain private.
34
+
35
+ ## Evaluation plan
36
 
37
+ Report held-out compiler acceptance, requirement coverage, model repair success, unsupported claims, and latency on the 5080 system. Keep all held-out projects out of training and prompt construction. Publish metrics and the exact training recipe with any released weights.
38
 
39
+ ## Release
40
 
41
+ Keep model weights private until proprietary-data leakage, licensing, compiler quality, and evaluation are reviewed. Studio discovers Ollama models installed on its serving host; the intended deployment name is mbes-sysml-qwen3-32b-v1. That tag must not be advertised as trained until the adapter is trained, tested, and installed.