Team Ai
Modelpublic

yassinsiouda/minimind-fr-electronics

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face
0likes39downloads
Model Card

minimind-fr-electronics

Electronics-repair q&a specialist. ~64 M params (hidden_size=768, num_hidden_layers=8, dense), MiniMind architecture. SFT from minimind-fr-electronics's base (pretrain-enfr -> base SFT -> agentic SFT).

Round-1 prototype — fluent but small; treat outputs accordingly. Training framework: <https://github.com/jingyaogong/minimind>.

Files

fileformat
spec-electronics_768.pthraw PyTorch state_dict (fp16) — load with MiniMindForCausalLM(strict=False)
spec-electronics-f16.ggufGGUF F16 — exported via Qwen3ForCausalLM, runs in llama.cpp / Ollama / LM Studio
tokenizer.json, tokenizer_config.jsonbyte-level BPE, vocab_size=6400, EN/FR

Config: num_attention_heads 8, num_key_value_heads 4, vocab_size 6400, max_position_embeddings 32768, rope_theta 1e6, tied embeddings, no MoE.

Training data

datasetcontribution
[`yassinsiouda/minimind-fr-electronics-data`](https://huggingface.co/datasets/yassinsiouda/minimind-fr-electronics-data)packaged training file for this model
`theprint/Electronics-QA`electronics Q&A (2,516)
`bshada/electronics.stackexchange.com`accepted answers, HTML stripped (30,000)
`allenai/tulu-3-sft-mixture`base-SFT replay (anti-forgetting)
`jpacifico/French-Alpaca-dataset-Instruct-110K`base-SFT replay
`angeluriot/french_instruct`base-SFT replay
`NousResearch/hermes-function-calling-v1`agentic replay
`nvidia/Nemotron-SFT-SWE-v3.5`terminal-loop replay

42,537 rows. English domain data, French carried by the bilingual base. ~25% replay of the base SFT mix; ~30% of rows carry a diagnostic <think> (symptom -> cause -> check). Recipe: convert_spec_electronics.py.

Run (GGUF)

bash
llama-cli -m spec-electronics-f16.gguf -p "..." -ngl 99

Limitations

64 M parameters; English domain data for the specialists (French comes from the bilingual base, so domain idiom is anglicised); agentic capability is SFT-only (short 2-4 step tool loops); safety limited to the router's thin unsafe bucket.

License

Apache-2.0 (weights). Upstream dataset licenses govern downstream use — see the dataset repo.