mtlm-7m-tools — a 7M-parameter assistant that talks English and calls tools, trained end-to-end in machin

Fine-tune of javimosch/mtlm-7m-base that decides between answering in plain English and emitting a one-line JSON tool call, over 14 generic tools (weather, calculator, web search, time, read/write file, HTTP GET, email, translate, shell, unit conversion, reminders, notes, wikipedia). Served by anvil, it speaks the OpenAI chat-completions protocol including tool_calls.

Everything except the synthesis of the fine-tuning text is pure machin (MFL): the base pretraining, the chat-template tokenizer, the fine-tune, the int8 export and the server. The text corpus was produced by a templated Python script from the tool catalog (no LLM teacher).

Honest numbers

base	mtlm-7m-base (7.2M params, TinyStories, val loss 1.916)
fine-tune data	40k synthetic conversations (76% plain-English assistant turns, 24% tool calls), 6.2M tokens, plus 15% TinyStories mixed into every batch
schedule	2,000 steps × 32 × 256, lr 2e-4, warmup 50, cosine; 6.2 h on a shared 6-core CPU
held-out loss	0.222 (step 100) → 0.088 (step 2,000)

Probes on the exported int8 model (temperature 0.3), probes.txt has the raw output:

prompt	output
what is the weather in Lyon?	{"tool_call":{"name":"get_weather","arguments":{"city":"Lyon"}}}
hello, who are you?	Hi! What can I do for you?
what is 123 plus 456?	{"tool_call":{"name":"calculator","arguments":{"expression":"123 + 456"}}}
send an email to lily@example.com with subject Hello saying See you at 10.	send_email with to, subject, body all correct
tell me a very short story about a cat	a fluent five-sentence story (about a girl, not a cat)
how do you say good morning in Spanish?	{"tool_call":{"name":"translate","arguments":{"text":"good morning","to":"Spanish"}}}
(tool result: Lyon, 18 °C, cloudy)	"In Lyon it is cloudy with -4 degrees." — right form, wrong number

Known limits: numbers are sometimes misquoted when summarizing a tool result; the tool set is fixed at fine-tune time (the model does not read tool schemas from the prompt — that is the point: 256 tokens of context, tools baked in); only the 14 tools above; English only; no knowledge beyond children's stories.

Chat format

anvil's Zephyr-style template at the token level: <s> then for each message <|role|>\n{content} </s> \n, and <|assistant|>\n to start generation. Tool results are rendered as a user message Tool result: {...}. A tool call is a single line {"tool_call":{"name":...,"arguments":{...}}} followed by </s>.

Run it as an OpenAI-compatible server

ANVIL_TOOLS_INJECT=0 anvil-serve models/m7tool2.bin 8097     # tools are baked in; do not paste schemas into the prompt
curl -s localhost:8097/v1/chat/completions -d '{
  "model": "mtlm", "temperature": 0.2,
  "messages": [
    {"role":"system","content":"You are a helpful assistant. You can call tools. When a tool is needed, reply with only the JSON tool call. Otherwise answer in plain English."},
    {"role":"user","content":"what is the weather in Lyon?"}],
  "tools": [{"type":"function","function":{"name":"get_weather","parameters":{"type":"object","properties":{"city":{"type":"string"}}}}}]}'

returns "tool_calls":[{"type":"function","function":{"name":"get_weather","arguments":"{\"city\":\"Lyon\"}"}}] with finish_reason: "tool_calls"; the same request with "hello, who are you?" returns plain text. Use the system prompt above: it is the one the model was trained with.

Files

• m7tool2.mtlm — fp32 checkpoint (mtlm1 layout).
• m7tool2.bin — llama2.c v2 int8 export, group size 32 (the model width, 288, is not a multiple of 64), 8.0 MB.
• tokenizer.bin — llama2.c tokenizer format, 4096 pieces.
• train_log.jsonl, probes.txt — the run log and the raw probe/server outputs above.