LFM2.5-24B-A2B · Harbor · pi (LoRA)

A LoRA adapter for LiquidAI/LFM2.5-24B-A2B, trained with reinforcement learning to work as the model behind the pi coding agent on data-analysis tasks.

Each training rollout is a full pi session in a remote sandbox: the agent reads the task, runs shell and Python tools against the task's data files, and answers. A verifier then grades the answer. The adapter was trained on those grades with TRL's asynchronous GRPO trainer.

In short: this checkpoint solves tasks about as often as the base model, using 39% fewer tool calls.

Results

Evaluated with the pi harness on the first 100 tasks of FineEnvs/SmolDataEnvs-harbor-test, one attempt per task at temperature 0.8, paired against the base model on the same tasks.

	vs base model
pass@1	−4.2 points (within noise: about ±5 points at 100 tasks)
Tool calls, on tasks both solved	−39%

This adapter is the orange line (pi, lr 1e-5) at step 100:

Reading the plot:

Accuracy stays within noise of the base model for the whole run. It dips at the very end (step 110), which is why this card publishes step 100.
Efficiency is what training changed. Tool calls on solved tasks fall steadily, from −4% at step 10 to −39% at step
1. The run at a lower learning rate (blue, 3e-6) and the runs with other harnesses (green, yellow) barely move.

Pick this adapter if you want a cheaper, faster agent at about the same accuracy. If accuracy is all you care about, the base model is as good.

Use with pi

1. Serve the adapter with vLLM

vllm serve LiquidAI/LFM2.5-24B-A2B \
    --enable-lora --max-lora-rank 32 \
    --lora-modules lfm-harbor-pi=harness-hub/LFM2.5-24B-A2B-harbor-pi-lora \
    --enable-auto-tool-choice --tool-call-parser lfm2 --reasoning-parser qwen3 \
    --max-model-len 131072

This is the configuration used in training and evaluation. --tool-call-parser lfm2 is required: pi's tool calls are parsed from the model's output.

2. Point pi at it

Install pi (Node.js 22.19 or newer):

npm install -g --ignore-scripts @earendil-works/pi-coding-agent

Add the server to ~/.pi/agent/models.json:

{
  "providers": {
    "vllm": {
      "baseUrl": "http://localhost:8000/v1",
      "api": "openai-completions",
      "apiKey": "dummy",
      "models": [{ "id": "lfm-harbor-pi" }]
    }
  }
}

"api": "openai-completions" matters: without it pi uses the Responses API, which this setup was not trained or evaluated with.

Run it:

pi --provider vllm --model lfm-harbor-pi
pi --provider vllm --model lfm-harbor-pi --print "Which region had the highest revenue growth in sales.csv?"

Use with llama.cpp

A GGUF version of this adapter is in harness-hub/LFM2.5-24B-A2B-harbor-pi-lora-GGUF. It runs on top of LiquidAI's base GGUF:

hf download LiquidAI/LFM2.5-24B-A2B-GGUF LFM2.5-24B-A2B-Q8_0.gguf --local-dir .
hf download harness-hub/LFM2.5-24B-A2B-harbor-pi-lora-GGUF LFM2.5-24B-A2B-harbor-pi-lora-F16.gguf --local-dir .

llama-server -m LFM2.5-24B-A2B-Q8_0.gguf --lora LFM2.5-24B-A2B-harbor-pi-lora-F16.gguf \
    --jinja -c 65536 --port 8080 --alias lfm-harbor-pi

--jinja is required for tool calling. Then use the pi configuration above with "baseUrl": "http://localhost:8080/v1". The GGUF version has not been re-evaluated.

Training


Method	AsyncGRPO (TRL, experimental), LoRA
Environment	Harbor tasks served by OpenEnv, rollouts in HF sandboxes
Harness	pi
Tasks	first 1,000 tasks of FineEnvs/SmolDataEnvs-harbor-train
Reward	the task verifier's `correctness` (falls back to `reward`)
Steps	100 (this checkpoint)
Learning rate	1e-5
LoRA	r = 32, alpha = 64, all linear layers
Rollouts	8 per task, temperature 0.8, up to 17 model calls and 600 s per rollout, 16,384 tokens per model call
Batch	4 token-budgeted micro-batches per step
Hardware	1× NVIDIA B300, vLLM and the trainer sharing the GPU

The training script is examples/async_grpo_harbor in TRL.

Limitations

Small evaluation. 100 tasks put the noise on pass@1 at about ±5 points, so none of the accuracy differences here are significant.
One task family. Training and evaluation are both on SmolDataEnvs data-analysis tasks. Gains may not carry over to other kinds of coding work.
One harness. The adapter was trained inside pi's prompts and tools. With another agent, expect it to behave like the base model, or worse.

harness-hub/LFM2.5-24B-A2B-pi