pitchlab-models

The working model store of Pitch, an in-store shopping assistant that answers customer questions from a merchant's own catalog. It holds the LoRA adapters, merged models and GGUF builds made while training per-merchant assistants, plus the base GGUFs they were served on. This is an archive of experiments, not a curated release.

The six live demos at demos.chiragpatnaik.com do not load from this repo. They load their small LFM2 builds from naklitechie/pitch-demo-weights.

Merchant names in folder names (Headphone Zone / hz, Equippers, Decathlon, 111Skin, Great Jones, Nicobar) mark which demo a model was built for. They do not mean the merchant made, reviewed or endorses the model.

Layout

Folder	What is in it
adapters/	24 PEFT LoRA adapters, rank 32: final adapter_model.safetensors + configs and tokenizers. Their checkpoint-* folders hold configs and training state only; the checkpoint weights were lost before upload.
serve/*/adapter/	4 serving adapters (hz-4b, hz-12b, qwen3-4b, qwen3-8b), same format
merged/, serve/lfm2-230m/	configs and tokenizers only; the merged weights were lost before upload (rebuild: merge the adapter into LFM2.5-230M)

Not included: GGUF builds (base models and LoRA GGUFs) and merged weights. Base GGUFs are public downloads from Google and Qwen; LoRA GGUFs rebuild from the adapters with llama.cpp's convert_lora_to_gguf.py.

adapters/

Base model	Adapters
LiquidAI/LFM2.5-230M	lfm2-230m-foundation-only, -hz, -hz-decline, -hz-toolpack, -decathlon, -greatjones, -111skin, -bakeoff-a
LFM2.5-230M + an earlier adapter merged in ("stacked")	lfm2-230m-hz-stacked, -equippers-stacked, -equippers-realid, -equippers-esc, -equippers-esc-v2, -equippers-esc-v3, -equippers-combined, -headphonezone-escv2, -headphonezone-escv3-stock, -decathlon-escv2, -nicobar-escv2, -protocol-generic, -protocol-prod
google/gemma-4-12B-it-qat-q4_0-unquantized	gemma-12b-equippers-realid, gemma-12b-equippers-realid-L
Qwen/Qwen3-4B-Instruct-2507	qwen3-4b-hz-toolpack

LFM2 adapters target attention and MLP projections (q/k/v/out_proj, in_proj, w1/w2/w3); Gemma and Qwen adapters target attention only (q/k/v/o_proj). For a stacked adapter, adapter_config.json names the local path of the merged base it was trained on; rebuild that base by merging the named earlier adapter into LFM2.5-230M first.

Licences

Each file follows the licence of its base model:

• LFM2.5-230M and everything built on it: LFM Open License v1.0
• Gemma 4 E4B and 12B and everything built on them: Apache-2.0 (E4B, 12B)
• Qwen3-4B-Instruct-2507 and Qwen3-8B and everything built on them: Apache-2.0

No warranty. These are research artefacts; outputs were evaluated only for the demos they were built for.