Fix accuracy benchmark commands

README.md

@@ -106,23 +106,29 @@ scp -r LFM2-2.6B-a16w4_return_logits sima@<modalix-ip>:/media/nvme/llima/models/
 
 ### Modalix Backend Accuracy
 
-Run Modalix accuracy benchmarking from a host machine using this precompiled `--return_logits` artifact:
+Run the benchmark from a host machine. The positional `model_id` must be the source Hugging Face model because MOLE uses its tokenizer and configuration; `--board_model` selects this compiled artifact on Modalix.
 
 ```bash
-llima-benchmark accuracy LFM2-2.6B-a16w4_return_logits -b modalix -o <output_dir> --max_num_tokens <max_num_tokens> \
-    --board_ip <board_ip> --board_port <board_port> \
-    --board_model LFM2-2.6B-a16w4_return_logits --board_start_server \
-    --board_venv_path <venv_on_board>
+MODALIX_IP="<modalix-ip>"
+llima-benchmark accuracy LiquidAI/LFM2-2.6B \
+    --backend modalix \
+    --output ./results/modalix \
+    --board_ip "$MODALIX_IP" \
+    --board_model LFM2-2.6B-a16w4_return_logits \
+    --task hellaswag piqa triviaqa wikitext winogrande
 ```
 
-The `--board_model` path must already exist on the Modalix device and contain both `devkit/` and `elf_files/`.
+The compiled `--board_model` must already exist on Modalix and contain both `devkit/` and `elf_files/`. The CLI starts the benchmark server by default. Add `--board_venv_path /path/to/venv` when `llima` is installed in a non-default virtual environment, or `--no-board_start_server` when connecting to a server that is already running.
 
-### HF Backend Accuracy
+### Hugging Face Backend Accuracy
 
-Run the Hugging Face backend reference benchmark from the host machine:
+Run the same task suite against the source model on the host:
 
 ```bash
-llima-benchmark accuracy LiquidAI/LFM2-2.6B -b hf -o <output_dir>
+llima-benchmark accuracy LiquidAI/LFM2-2.6B \
+    --backend hf \
+    --output ./results/hf \
+    --task hellaswag piqa triviaqa wikitext winogrande
 ```
 
 See [MOLE accuracy benchmarking](https://developer.sima.ai/software/genai-llima/mole) for the full accuracy benchmarking workflow.