A 0.8B-parameter model that reads one English chat message and lists every clock time in it, deciding from context whether a bare "7" means morning or evening. It is the middle option of the project: more accurate on unfamiliar situations than the 230M model, a quarter of the size of the 2B model, and small enough to run in a browser tab. It is too slow for a 100 ms budget on CPU; for that, see https://huggingface.co/dsego/tempus-lfm2.5-230m
Code, data and the full results: https://github.com/dsegovia90/tempus_finetune
| Message | Reply |
|---|---|
dinner at 7 or lunch at noon |
7=19:00noon=12:00 |
got paged at 2, been up ever since |
2=02:00 |
skip to 2:30 in the video, room 3 is booked |
none |
| File | Size | Use |
|---|---|---|
qwen3.5-0.8b-v2-Q4_K_M.gguf |
542 MB | Recommended. Within one message of Q8_0 on all three evals |
qwen3.5-0.8b-v2-Q8_0.gguf |
834 MB | 8-bit weights |
qwen3.5-0.8b-v2-trim-Q4_K_M.gguf |
448 MB | The Q4_K_M model with its vocabulary cut to the tokens English chat needs, for a smaller browser download |
Text only: the base model's vision component is not included.
Exact match per message: every time found, with the right span and the right 24-hour reading. This is the best of three training runs, chosen by these same evals; the three-run mean is in the right-hand column.
| Eval | Q4_K_M | Q8_0 | Three-run mean (Q8_0) |
|---|---|---|---|
| Main (530 messages) | 97.0% | 97.2% | 96.7% |
| Main, time of day implied by the situation only | 91.5% | 92.1% | 90.6% |
| Fresh (378 messages from situations not in training) | 98.4% | 98.4% | 98.1% |
| Fresh, time of day implied by the situation only | 95.5% | 95.5% | 94.8% |
| Wrong replies on 1,021 tricky messages with no clock time | 7 (0.7%) | 8 (0.8%) | 0.6% |
| CPU latency, median, 4 cores of a shared Ryzen 3600 | 340 ms | 421 ms | 412 ms |
| GPU latency, median, RTX 3090 | 39 ms | 40 ms | 41 ms |
About half of the wrong replies on the no-time messages are not valid times ("35=35:00" for "it's 35 degrees out"), so a client that checks the reply, as the repository's does, discards them.
The trimmed file scores 97.0% on the main eval, with every reply identical to the untrimmed Q4_K_M file's, 98.1% on the fresh eval and 5 wrong replies (0.5%) on the no-time messages. It differs from the untrimmed file on 5 of the other 1,399 messages; those two runs were also on different processors (CPU and GPU), which flips near-ties by itself.
With llama.cpp (tested at tag b12292), start
llama-server --model qwen3.5-0.8b-v2-Q4_K_M.gguf and send the message as a single user turn
with no system prompt and thinking turned off. The raw prompt is:
<|im_start|>user
{message}<|im_end|>
<|im_start|>assistant
<think>
</think>
The reply is one text=HH:MM line per clock time, where text is copied from the message, or
the single word none. Use greedy decoding (temperature 0).
In a browser through llama.cpp's WebAssembly build, this model's saved prompt states are large and fill the memory after about 80 replies; the repository's demo restarts the worker every 16 replies to avoid that.
LoRA (rank 16) on Qwen/Qwen3.5-0.8B with Unsloth,
3 epochs over 11,299 synthetic messages, merged and converted to GGUF. Config:
configs/qwen3.5-0.8b-v2-s0-compact.yaml in the repository.
Apache 2.0, the same as the base model.
Qwen/Qwen3.5-0.8B