coglang-drug-distill

coglang-drug-distill — drug-v6 (Q4_K_M GGUF)

A CogLang-specialist small model for on-device inference. It translates natural-language medication-safety questions into CogLang — a graph-first intermediate language — which is then executed against a curated drug-interaction graph.

The model only writes graph queries; it never answers drug questions from memory. Pharma facts live in the graph, not the weights.

⚠️ Medical disclaimer. Learning/research demo only. Not a medical device, not clinical decision support. The accompanying graph is a tiny illustrative seed, not authoritative. Never use any output for real medical decisions.

Files

• Base model: Qwen/Qwen2.5-1.5B-Instruct
• Method: QLoRA (4-bit) fine-tune → merged → GGUF → Q4_K_M quantize
• Size / footprint: 935 MiB, ~1 GB RAM
• Speed: ~33 tok/s gen on CPU (i9-13900HX, 8 threads)
• Full code & pipeline: https://github.com/zhpy2004/coglang-drug-distill

Results

drug-v6 bakes the SYSTEM_v6d verb→operator rules into the weights and adds targeted training data for the operators v5 still missed (Traverse-as-"causes", ForEach, Trace[...] wrapping, Do[...] pairing). This cured those routing errors — OOD is now saturated at 100%. The remaining ~3 wild misses are structural limits of single-shot CogLang (multi-relation union, reverse traversal), not routing.

Pairing the model with the CogLang GBNF grammar guarantees 100% parseable output, so the model only has to learn semantics, not syntax.

Usage (llama.cpp, grammar-constrained)

# get the CogLang GBNF grammar from the upstream repo:
#   https://github.com/jaysinailabs/coglang  ->  examples/grammar/coglang.gbnf

llama-cli -m drug-v6-q4_k_m.gguf \
  -st \
  -sys "$(cat SYSTEM_v6d.txt)" \
  -p "warfarin interactions" \
  --grammar-file coglang.gbnf \
  -n 256 --temp 0 -t 8
# -> Traverse["warfarin", "interacts_with"]

The tuned system prompt (SYSTEM_v6d.txt) and few-shot blocks live in the GitHub repo. An interactive CLI (pipeline/drug_chat.py) and a modular local web app (app/, runnable on desktop and in Termux on a phone) wire NL → CogLang → execution against the seed graph.

License & attribution

Apache-2.0. Built on and depends on CogLang (Jason Xin / Bigcoders, Apache-2.0), which is not redistributed here.