JSL-Granite-CP API & Inference Endpoint

Results — clean held-out test (700 rows, zero training leakage)

Table with columns: model, F1, hallucinations
model	F1	hallucinations
base `granite-4.0-h-tiny`	88.33	288
JSL-Granite-CP (this model)	93.97	39

Scored with the project's shipped reward logic (grpo_reward / offline simulate_cp — the rule is executed against the text and matched spans are compared to gold). Test rows share 0 examples with the training data.

Training

QLoRA (bitsandbytes 4-bit nf4, LoRA r=16 / α=32, target_modules="all-linear") on 5,029 reward-verified CP-rule examples (balanced across entities, weak entities oversampled), 2 epochs, then merged back to bf16.

Usage

Use the project system prompt (from to_chat.py). Serve with vLLM:

bash

vllm serve Zaynoid/JSL-Granite-CP --max-model-len 8192 --trust-remote-code

User message format:

text

Text: <clinical text>Target entity: <ENTITY>Return the ContextualParser rule as JSON.

Output: {"cp_rule": { ... }} or {"cp_rule": null}.

Results — clean held-out test (700 rows, zero training leakage)

Table with columns: model, F1, hallucinations
model	F1	hallucinations
base `granite-4.0-h-tiny`	88.33	288
JSL-Granite-CP (this model)	93.97	39

Training

Usage

Use the project system prompt (from to_chat.py). Serve with vLLM:

bash

vllm serve Zaynoid/JSL-Granite-CP --max-model-len 8192 --trust-remote-code

User message format:

text

Text: <clinical text>Target entity: <ENTITY>Return the ContextualParser rule as JSON.

Output: {"cp_rule": { ... }} or {"cp_rule": null}.

JSL-Granite-CP

README

Results — clean held-out test (700 rows, zero training leakage)

Training

Usage

Explore FriendliAI today

README

Results — clean held-out test (700 rows, zero training leakage)

Training

Usage