Add model card (negative result)

README.md

@@ -0,0 +1,122 @@
+---
+license: gemma
+base_model: lmstudio-community/gemma-4-E4B-it-MLX-4bit
+tags:
+- mlx
+- lora
+- peft
+- electronics
+- embedded
+- kicad
+- spice
+- ailiance
+- negative-result
+language:
+- en
+- fr
+pipeline_tag: text-generation
+---
+
+# Ailiance — gemma-4-E4B kicad9plus LoRA
+
+> ⚠️ **NEGATIVE RESULT — PUBLISHED FOR TRANSPARENCY**
+>
+> This LoRA exhibits **catastrophic forgetting on SPICE tasks** (-15.7 pts composite vs base)
+> and **complete loss of schematic extraction capability** (P3 = 0.000, -30.8 pts vs base).
+> Published as a research artifact to document the failure mode, not for production use.
+>
+> **Hypothesis**: 98 training samples on `kicad9plus-permissive` + rank=8 / scale=20.0 default
+> were insufficient for grammar-style learning, and the high `scale=20.0` over-amplified the
+> small rank-8 deltas onto the base model, breaking unrelated capabilities.
+>
+> **Future iterations may revisit** with rank=32, lower scale, and a larger corpus
+> (e.g. `Ailiance-fr/kicad9plus-copyleft` adds 209 GPL samples for 307 total).
+
+LoRA adapter fine-tuned on `lmstudio-community/gemma-4-E4B-it-MLX-4bit` for KiCad 9+ schematic-file (`.kicad_sch`) generation. **Did not converge to useful behaviour.**
+
+> Maintained by **Ailiance** — French AI org building EU AI Act compliant resources for embedded systems and electronics design.
+
+## Quick start (for reproduction)
+
+```python
+from mlx_lm import load, generate
+
+model, tokenizer = load(
+    "lmstudio-community/gemma-4-E4B-it-MLX-4bit",
+    adapter_path="Ailiance-fr/gemma-4-E4B-kicad9plus-lora"
+)
+# Expect degraded behaviour on SPICE & extraction tasks.
+```
+
+## Training
+
+- **Base model**: `lmstudio-community/gemma-4-E4B-it-MLX-4bit`
+- **Training data**: [`Ailiance-fr/kicad9plus-permissive`](https://huggingface.co/datasets/Ailiance-fr/kicad9plus-permissive) — 98 KiCad 9+ schematic samples under permissive licenses (CC-BY-SA-4.0)
+- **Method**: LoRA (PEFT) via `mlx-lm`
+- **Iterations**: 1200
+- **Rank**: 8 *(low — suspected root cause)*
+- **Scale**: 20.0 *(high — suspected root cause)*
+- **Dropout**: 0.0
+- **Learning rate**: 1e-5
+- **Max seq length**: 4096
+- **Layers**: 16 (partial — not all)
+- **Curriculum**: none — single dataset, single pass.
+
+The `rank=8 / scale=20.0` combination is the `mlx-lm` lora default and was kept unchanged for this experiment. With only 98 samples, rank 8 likely undersampled the schematic grammar's complexity while the high scale propagated noisy deltas into unrelated capability channels — hence the broken SPICE + extraction scores.
+
+## Benchmark results
+
+Composite scores from `electron-bench` (see [compare_base_vs_lora.md](https://github.com/ailiance/ailiance-bench/blob/main/bench-results/compare_base_vs_lora.md)). Reference base = `gemma-4-E4B-it-MLX-4bit`.
+
+| Phase | Dataset            |  base | this LoRA |       Lift |
+|------:|--------------------|------:|----------:|-----------:|
+| P1    | kicad-dsl          | 0.090 |     0.090 |     +0.0pts |
+| P1    | kicad-pcb          | 0.010 |     0.015 |     +0.5pts |
+| P1    | spice-sim          | 0.425 |     0.268 | **-15.7pts** |
+| P2    | kicad-sch-gen      | 0.420 |     0.180 | **-24.0pts** |
+| P3    | kicad-sch-extract  | 0.308 |     0.000 | **-30.8pts** |
+| P4    | kicad-erc-abs      | 0.060 |     0.033 |    -2.7pts |
+| P5    | kicad-erc-delta    | 0.060 |     0.033 |    -2.7pts |
+
+**Verdict**: clear catastrophic forgetting. P3 collapses to zero (the model can no longer produce parseable extraction output), and the *very task this LoRA was trained for* — `kicad-sch-gen` (P2) — actually got **worse** (-24 pts). Published openly so the community can learn from the failure mode rather than rediscover it.
+
+## What we learned
+
+- 98 samples is below the practical floor for grammar-style learning on this base model.
+- `rank=8 + scale=20.0` is a dangerous default for narrow corpora; the high scale acts like a learning-rate multiplier and corrupts capabilities outside the training distribution.
+- Future runs should use `rank=32, scale=2.0` (same as `eukiki` / `mascarade`) plus a larger corpus.
+
+## EU AI Act compliance
+
+- **Article 53(1)(c) copyright policy**: training data licenses preserved (CC-BY-SA-4.0 from upstream `Ailiance-fr/kicad9plus-permissive`).
+- **Article 53(1)(d) training data summary**:
+  - Publicly available datasets: `Ailiance-fr/kicad9plus-permissive` (98 samples, CC-BY-SA-4.0).
+  - Web scraping: No.
+  - Licensed data: None.
+- **GPAI Code of Practice (July 2025)**: base model Gemma (Google = signatory).
+- **Provenance**: per-sample `metadata.license_spdx` preserved in upstream dataset.
+
+## License
+
+- LoRA adapter weights: released under **Gemma Terms of Use** (inheritance from base model).
+- See https://ai.google.dev/gemma/terms
+
+## Citation
+
+```bibtex
+@misc{ailiance_gemma4_e4b_kicad9plus_lora_2026,
+  author    = {Ailiance},
+  title     = {Ailiance — gemma-4-E4B kicad9plus LoRA (negative result)},
+  year      = {2026},
+  publisher = {Hugging Face},
+  url       = {https://huggingface.co/Ailiance-fr/gemma-4-E4B-kicad9plus-lora},
+  note      = {Published for transparency; catastrophic forgetting on SPICE and P3 extraction.}
+}
+```
+
+## Related models
+
+- [Ailiance-fr/gemma-4-E4B-eukiki-lora](https://huggingface.co/Ailiance-fr/gemma-4-E4B-eukiki-lora) — champion general
+- [Ailiance-fr/gemma-4-E4B-mascarade-lora](https://huggingface.co/Ailiance-fr/gemma-4-E4B-mascarade-lora) — champion schematic extraction
+- [Ailiance-fr/gemma-4-E4B-aggro-test-lora](https://huggingface.co/Ailiance-fr/gemma-4-E4B-aggro-test-lora) — sanity-check baseline
+- [Ailiance-fr/gemma-4-E4B-kicad9plus-lora](https://huggingface.co/Ailiance-fr/gemma-4-E4B-kicad9plus-lora) — negative result (this model, published for transparency)