初始化项目，由ModelHub XC社区提供模型

Model: VittoriaLanzo/Ohmatic-Qwen3-8B Source: Original Platform
2026-07-07 20:09:38 +08:00
commit b7657d7add
19 changed files with 152691 additions and 0 deletions
--- a/README.md
+++ b/README.md
@@ -0,0 +1,88 @@
+---
+license: other
+license_name: ohmatic-sal-1.1
+license_link: LICENSE
+base_model: Qwen/Qwen3-8B
+language:
+- en
+pipeline_tag: text-generation
+tags:
+- circuit-design
+- schematic-generation
+- electronics
+- erc
+- qwen3
+- self-correction
+- gguf
+---
+
+# Ohmatic-Qwen3-8B
+
+**Ohmatic** generates electronic circuit schematics from natural-language descriptions and
+*corrects its own designs* against an Electrical Rule Checker (ERC). It is a fully assembled
+(merged, no adapter required) 8B model based on Qwen3-8B.
+
+## How it works
+
+Ohmatic is trained to operate as a closed verification loop, not a one-shot generator:
+
+1. **Forward generation** - the user describes a circuit in plain language; the model emits a
+   complete structured schematic (components, values, nets).
+2. **ERC verification** - the schematic is checked by a deterministic Electrical Rule Checker
+   (shorts, floating nets, missing references, polarity/supply errors, unclosed structures).
+3. **Self-correction** - on ERC failure, the model receives the rule-checker findings and emits a
+   *repaired* schematic. Training explicitly teaches this correction turn, so the model improves
+   designs rather than re-rolling them.
+
+## Training
+
+Trained to both produce circuits and repair its own designs from ERC feedback, using only
+ERC-verified examples. The released weights are **fully merged** - load like any causal LM, no
+PEFT/adapter required.
+
+- **Base**: Qwen3-8B (bf16)
+- The training data, recipe, and ERC engine are **proprietary**; this card documents the model
+  artifact you run.
+
+## Files
+
+| File | Format | Use |
+|---|---|---|
+| `*.safetensors` | bf16, sharded | transformers / vLLM serving, further finetuning |
+| `Ohmatic-Qwen3-8B-Q8_0.gguf` | GGUF 8-bit | llama.cpp / LM Studio / ollama - near-lossless |
+| `Ohmatic-Qwen3-8B-Q4_K_M.gguf` | GGUF 4-bit | llama.cpp on consumer hardware |
+
+## Usage (transformers)
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+m  = AutoModelForCausalLM.from_pretrained("VittoriaLanzo/Ohmatic-Qwen3-8B",
+                                          torch_dtype="bfloat16", device_map="auto")
+tk = AutoTokenizer.from_pretrained("VittoriaLanzo/Ohmatic-Qwen3-8B")
+msgs = [{"role": "user", "content": "Design a 5V-to-3.3V LDO supply with input protection."}]
+x = tk.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(m.device)
+print(tk.decode(m.generate(x, max_new_tokens=2048)[0], skip_special_tokens=True))
+```
+
+## Usage (llama.cpp)
+
+```bash
+llama-cli -m Ohmatic-Qwen3-8B-Q4_K_M.gguf -cnv \
+  -p "Design an astable 555 timer blinking an LED at 1 Hz on 9V."
+```
+
+## Evaluation
+
+Held-out ERC pass rate at selection time (n=32 in-training eval): **53.1% first-pass** validity,
+with the correction loop recovering a further share of failures. This is the single-shot
+held-out number; the full **product-pipeline** benchmark (normalization + correction loop +
+killswitch, judged by the same ERC engine) is reported in the
+[Ohmatic repository](https://github.com/VittoriaLanzo/Ohmatic#benchmark).
+
+## License
+
+**Ohmatic Source-Available License 1.1 (Ohmatic-SAL-1.1)** - adapted from the Functional Source
+License 1.1, but it is **not** the FSL: the only change is a 10-year change date (instead of two),
+after which the grant converts to Apache-2.0. Full text in [LICENSE](LICENSE). Source-available,
+not open source: any Permitted Purpose is allowed, a Competing Use is not. (Base model
+`Qwen/Qwen3-8B` is separately licensed; these merged weights are Ohmatic-SAL-1.1.)