From 6c8cce620e3d7039df95c4f05872cab4ec853bc8 Mon Sep 17 00:00:00 2001
From: Kasmesq <Kasmesq@users.noreply.github.com>
Date: Wed, 8 Jul 2026 22:42:28 -0400
Subject: [PATCH] Add FastKV LongBench Llama run wrapper

---
 .../evaluation/FASTKV_LONGBENCH_LLAMA31.md    | 58 +++++++++++++++++++
 .../run_fastkv_longbench_llama31_8b_20_30.sh  | 10 ++++
 2 files changed, 68 insertions(+)
 create mode 100644 experiments/llmlingua2/evaluation/FASTKV_LONGBENCH_LLAMA31.md
 create mode 100755 experiments/llmlingua2/evaluation/scripts/run_fastkv_longbench_llama31_8b_20_30.sh

diff --git a/experiments/llmlingua2/evaluation/FASTKV_LONGBENCH_LLAMA31.md b/experiments/llmlingua2/evaluation/FASTKV_LONGBENCH_LLAMA31.md
new file mode 100644
index 0000000..310b331
--- /dev/null
+++ b/experiments/llmlingua2/evaluation/FASTKV_LONGBENCH_LLAMA31.md
@@ -0,0 +1,58 @@
+# FastKV LongBench + LLMLingua-2 + Llama 3.1 8B Instruct
+
+This setup runs LLMLingua-2 compression on the LongBench data stored in a sibling
+FastKV checkout, then evaluates with `meta-llama/Llama-3.1-8B-Instruct` at
+adaptive 20% and 30% prompt-token budgets.
+
+Expected checkout layout:
+
+```text
+<workdir>/FastKV
+<workdir>/LLMLingua
+```
+
+Required local data and prompt source:
+
+```text
+FastKV/data/LongBench/*.jsonl
+FastKV/eval/run_longbench.py
+```
+
+Before running, authenticate to Hugging Face with an account that has access to
+Llama 3.1 8B Instruct. Do not commit tokens.
+
+```bash
+huggingface-cli login
+```
+
+Run from either repository:
+
+```bash
+cd FastKV
+bash scripts/run_llmlingua2_llama31_8b_20_30.sh
+```
+
+or
+
+```bash
+cd LLMLingua
+bash experiments/llmlingua2/evaluation/scripts/run_fastkv_longbench_llama31_8b_20_30.sh
+```
+
+Useful overrides:
+
+```bash
+FASTKV_ROOT=/path/to/FastKV LLMLINGUA_REPO=/path/to/LLMLingua DATA_DIR=/path/to/LongBench SAVE_DIR=/path/to/results RATES="0.20 0.30" DATASETS="narrativeqa,qasper" MAX_NUM_EXAMPLES=10 bash scripts/run_llmlingua2_llama31_8b_20_30.sh
+```
+
+Default output:
+
+```text
+LLMLingua/results/fastkv_longbench_llmlingua2/llama3_1_8b_instruct/adaptive20
+LLMLingua/results/fastkv_longbench_llmlingua2/llama3_1_8b_instruct/adaptive30
+```
+
+The adaptive budget is computed per sample as a percentage of the full downstream
+prompt length after applying the FastKV LongBench prompt and chat template. The
+runner subtracts prompt overhead and passes the remaining target token count to
+LLMLingua-2 for context compression.
diff --git a/experiments/llmlingua2/evaluation/scripts/run_fastkv_longbench_llama31_8b_20_30.sh b/experiments/llmlingua2/evaluation/scripts/run_fastkv_longbench_llama31_8b_20_30.sh
new file mode 100755
index 0000000..5fb84c9
--- /dev/null
+++ b/experiments/llmlingua2/evaluation/scripts/run_fastkv_longbench_llama31_8b_20_30.sh
@@ -0,0 +1,10 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
+LLMLINGUA_REPO=${LLMLINGUA_REPO:-$(cd "$SCRIPT_DIR/../../../.." && pwd)}
+FASTKV_ROOT=${FASTKV_ROOT:-$(cd "$LLMLINGUA_REPO/.." && pwd)/FastKV}
+
+export LLMLINGUA_REPO FASTKV_ROOT
+
+exec bash "$FASTKV_ROOT/scripts/run_llmlingua2_llama31_8b_20_30.sh"