Add v9 + summaries column; recommend running v9 bare (no --llm-summaries)

Update card with v3.v2 production-eval results + un-retirement note

Update model card with corrected eval numbers (296q expanded eval, CoIR 19-subtask avg)

Upload README.md with huggingface_hub

Upload tokenizer_config.json with huggingface_hub

Upload config.json with huggingface_hub

Upload README.md with huggingface_hub

docs: add full CoIR 9-task breakdown to model card

docs: fix model card accuracy — controlled base numbers, mean pooling example, remove stale hard eval claim

docs: add The Stack to datasets (v7 trains on CSN + Stack Rust/TS/C++)

eval: fix dataset IDs to CoIR-Retrieval org + per-language CSN scores

eval: add CoIR v7 results (9 tasks, NDCG@10)

docs: update model card for LoRA v7 (0.707 CSN, 49.19 CoIR)

feat: LoRA v7 ONNX opset 11 (compatible with ort CUDA EP)

fix: add shape inference (997 value_info entries) for CUDA EP compatibility

revert: restore LoRA v5 ONNX (v7 has ORT compatibility issue)

feat: LoRA v7 — update tokenizer_config.json

feat: LoRA v7 — GIST+Matryoshka+hard negs (0.707 CSN, 49.19 CoIR)

Fix: training hardware is A6000, not RTX 4090

Model card updated for LoRA v5 (166k/1ep, 0.683 CSN, 0.348 CosQA)

Switch to LoRA v5 (166k/1ep): +1.2pp CSN, +1.4pp CosQA vs v3

Switch to LoRA v5 (166k/1ep): +1.2pp CSN, +1.4pp CosQA vs v3

Update model card: v5 incoming, corrected metrics table

Add model card with benchmarks, training details, and usage examples

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

initial commit