XT-Luna-1B

XT-Luna-1B is a full fine-tune (all 1.08B parameters) of openbmb/MiniCPM5-1B, built as a tool-using assistant with an emphasis on restraint and honesty:

• declines requests whose capability is not in the provided tool list (instead of inventing tools),
• asks for explicit confirmation before destructive actions,
• batches independent tool calls in a single turn,
• reports tool failures truthfully (no fake success),
• plus ordinary single-call tool use, chit-chat without spurious calls, clarifying questions for ambiguous requests, and code generation.

It supports both decode modes of the base template: thinking (enable_thinking=True) and non-thinking (enable_thinking=False), and detects the active tool list from the prompt (trained on a 16-tool phone-assistant surface and a coding-tool surface).

Evaluations

At a glance

benchmark	XT-Luna-1B	MiniCPM5-1B-SFT	Qwen3.5-4B
xt_amelia (sum of 16, higher better)	11.25	7.00	9.50
code_eval pass@1	0.85 (17/20)	0.70 (14/20)	0.60 (12/20)
honesty_probe overall	0.72	0.33	0.44

xt_amelia suite

19 probes; family score = pass fraction (0.25 granularity). n = non-thinking, t = thinking.

family	XT-Luna n	XT-Luna t	SFT n	SFT t	Qwen n	Qwen t
single_call_exact	0.75	0.50	0.75	0.75	0.00	0.00
multi_tool	0.00	0.50	0.00	0.50	0.00	0.00
destructive_confirm	0.00	0.50	0.00	0.00	0.50	0.50
unavailable_tool	0.50	1.00	1.00	0.50	1.00	1.00
tool_failure_honesty	1.00	1.00	1.00	1.00	1.00	1.00
chit_chat_no_call	1.00	1.00	0.50	1.00	1.00	1.00
ambiguous_clarify	0.50	1.00	0.00	0.00	0.00	0.50
code_c	1.00	1.00	0.00	0.00	1.00	1.00
sum (of 16)	4.75	6.50	3.25	3.75	4.50	5.00

code_eval (pass@1, 20 tasks)

group	XT-Luna-1B	SFT	Qwen3.5-4B
Python non-thinking (8)	8/8	7/8	8/8
Python thinking (8)	8/8	7/8	0/8
C non-thinking (2)	1/2	0/2	2/2
C thinking (2)	0/2	0/2	2/2
total	17/20 (0.85)	14/20 (0.70)	12/20 (0.60)

Notable: Qwen3.5-4B's Python collapses in thinking mode (0/8) while its C generation is clean; XT-Luna is the reverse — strong Python in both modes, weak C (only c_sum non-thinking passes).

honesty_probe (6 categories)

category	XT-Luna-1B	SFT	Qwen3.5-4B
fake_success	1.00	1.00	1.00
done_when_broken	1.00	1.00	1.00
sandbagging	1.00	0.00	0.00
promise	0.67	0.00	0.00
identity	0.00	0.00	0.00
calibration	0.67	0.00	0.67
overall	0.72	0.33	0.44

Caveats

The probe suites are small (xt_amelia: 19 probes; code_eval: 20 tasks; honesty_probe: 6 categories) and the model was tuned on data rendered in a matching chat/tool format; comparisons to other models are indicative, not a general capability ranking.

Known limitations

• multi_tool in non-thinking mode: 0.00 — batching independent calls works in thinking mode only.
• destructive_confirm in non-thinking mode: 0.00 — confirmation behavior is reliably elicited only with thinking enabled.
• unavailable_tool non-thinking: 0.50 — sometimes calls a plausible but absent tool.
• identity probe: 0/3 across all tested models (identity questions are not handled reliably).
• C code generation is weak (1/4); keep C tasks in Python or double-check.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("XtrCom/XT-Luna-1B")
model = AutoModelForCausalLM.from_pretrained("XtrCom/XT-Luna-1B", dtype="bfloat16")

tools = [
    {"type": "function", "function": {
        "name": "set_alarm",
        "description": "Set an alarm.",
        "parameters": {"type": "object",
                       "properties": {"label": {"type": "string"},
                                      "time": {"type": "string"}},
                       "required": ["label", "time"]}}},
]

messages = [{"role": "user", "content": "Set a gym alarm for 6:30."}]
prompt = tok.apply_chat_template(
    messages, tools=tools, tokenize=False,
    add_generation_prompt=True, enable_thinking=False,
)
ids = tok(prompt, return_tensors="pt").input_ids
out = model.generate(ids, max_new_tokens=512, do_sample=False)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=False))

Tool calls are emitted as <function name="..."><param name="...">...</param></function> blocks. When decoding, do not skip special tokens if you need to parse call boundaries (<|im_end|>).

GGUF quantizations (Q5_K_M, F16) are available in XtrCom/XT-Luna-1B-GGUF.

Training

Staged full fine-tuning on a single RTX 3090 (bf16, maxlen 2048, lr 3e-6, 480-step final stage, selected by checkpoint sweep):

1. curriculum corpus of agentic tool-use, honesty, prose, and code rows,
2. targeted restraint data (declines, confirmations, clarifications, batching, failure honesty) rendered through the same chat template,
3. final balanced mixture of corpus anchors + restraint rows on both a coding-tool surface and a 16-tool phone-assistant surface.

Training data includes synthetic rows generated with a commercial LLM API and hand-curated examples. Intermediate checkpoints are archived at XtrCom/XT-Luna-1B-checkpoints.

License

AGPL-3.0 (see LICENSE). This model is a derivative of openbmb/MiniCPM5-1B; the base model's license and terms also apply to the base weights.

Released 2026-10-17 by XtrCom.