functiongemma-270m-it-intercomswap-v3

IntercomSwap fine-tuned FunctionGemma model for deterministic tool-calling in BTC Lightning <-> USDT Solana swap workflows.

What Is IntercomSwap

Intercom Swap is a fork of upstream Intercom that keeps the Intercom stack intact and adds a non-custodial swap harness for BTC over Lightning <> USDT on Solana via a shared escrow program, with deterministic operator tooling, recovery, and unattended end-to-end tests.

GitHub: https://github.com/TracSystems/intercom-swap

Base model: google/functiongemma-270m-it

Model Purpose

• Convert natural-language operator prompts into validated tool calls.
• Enforce buy/sell direction mapping for swap intents.
• Support repeat/autopost workflows used by IntercomSwap prompt routing.

Repository Layout

• ./:
  • merged HF checkpoint (Transformers format)
• ./nvfp4:
  • NVFP4-quantized checkpoint for TensorRT-LLM serving
• ./gguf:
  • functiongemma-v3-f16.gguf
  • functiongemma-v3-q8_0.gguf

Startup By Flavor

1) Base HF checkpoint (Transformers)

python -m vllm.entrypoints.openai.api_server \
  --model TracNetwork/functiongemma-270m-it-intercomswap-v3 \
  --host 0.0.0.0 \
  --port 8000 \
  --dtype auto \
  --max-model-len 8192

Lower memory mode example:

python -m vllm.entrypoints.openai.api_server \
  --model TracNetwork/functiongemma-270m-it-intercomswap-v3 \
  --host 0.0.0.0 \
  --port 8000 \
  --dtype auto \
  --max-model-len 4096 \
  --max-num-seqs 8

2) NVFP4 checkpoint (./nvfp4)

TensorRT-LLM example with explicit headroom (avoid consuming all VRAM):

trtllm-serve serve ./nvfp4 \
  --backend pytorch \
  --host 0.0.0.0 \
  --port 8012 \
  --max_batch_size 8 \
  --max_num_tokens 16384 \
  --kv_cache_free_gpu_memory_fraction 0.05

Memory tuning guidance:

• Decrease --max_num_tokens first.
• Then reduce --max_batch_size.
• Keep --kv_cache_free_gpu_memory_fraction around 0.05 to preserve safety headroom.

3) GGUF checkpoint (./gguf)

Q8_0 (recommended default balance):

llama-server \
  -m ./gguf/functiongemma-v3-q8_0.gguf \
  --host 0.0.0.0 \
  --port 8014 \
  --ctx-size 8192 \
  --batch-size 256 \
  --ubatch-size 64 \
  --gpu-layers 12

F16 (higher quality, higher memory):

llama-server \
  -m ./gguf/functiongemma-v3-f16.gguf \
  --host 0.0.0.0 \
  --port 8014 \
  --ctx-size 8192 \
  --batch-size 256 \
  --ubatch-size 64 \
  --gpu-layers 12

Memory tuning guidance:

• Lower --gpu-layers to reduce VRAM usage.
• Lower --ctx-size to reduce RAM+VRAM KV-cache usage.
• Use q8_0 for general deployment, f16 for quality-first offline tests.

Training Snapshot

• Base family: FunctionGemma 270M instruction-tuned.
• Fine-tune objective: IntercomSwap tool-call routing and argument shaping.
• Corpus profile: operations + intent-routing + tool-calling examples.

Evaluation Snapshot

From held-out evaluation for this release line:

• Train examples: 6263
• Eval examples: 755
• Train loss: 0.01348
• Eval loss: 0.02012

Intended Use

• Local or private deployments where tool execution is validated server-side.
• Deterministic operator workflows for swap infra.

Out-of-Scope Use

• Autonomous financial decision-making.
• Direct execution of unvalidated user text as shell/actions.
• Safety-critical usage without host-side policy/validation.

Safety Notes

• Always validate tool name + argument schema server-side.
• Treat network-side payloads as untrusted input.
• Keep wallet secrets and API credentials outside model context.

Provenance

• Derived from: google/functiongemma-270m-it
• Integration target: IntercomSwap prompt-mode tool routing