Spaces:

---

StentorLabs

/

SLM_Arena

Like 23

Running

Hi, I'd like to add peacebell-v1-148M to the Instruct models: https://huggingface.co/wayneworkman2012/peacebell-v1-148M

148.55M parameters, trained from scratch, Apache-2.0, ChatML chat model. It is a single-domain model: it only knows World War II and redirects most other topics back to the war. It is a first version and I would like it out there next to the others.

I know you normally take requests as a discussion and add the row yourself. I opened a PR because a row alone would not work for this model, and I wanted to hand you something already tested. Feel free to apply it your own way instead of merging.

Why a row alone is not enough

I ran this Space's own code (everything above the UI section of app.py) against the model locally, on CPU. It loads fine through _load_tokenizer / _load_model, but every reply came back as [Generation Error] 'GPTForCausalLM' object has no attribute 'generate'. Two reasons, both by design in my repo:

• the model ships its own KV-cached decoder (stream_generate) instead of HF generate();
• it has no Jinja chat_template: its ChatML control tokens are injected by id exactly as at training time (tokenizer.build_chatml_ids), which a string template re-tokenized by SentencePiece cannot reproduce.

What this PR changes (app.py: 29 lines added, 0 removed or edited; README.md: 1 line added)

• one row in INSTRUCT_MODEL_ROWS and one line in the README models: list;
• _is_peacebell_spec() and _stream_peacebell_text(), plus a three-line branch in _stream_single_text after the shared model load and stop-signal reset. It yields the same cumulative partial text as _stream_text_from_model, honours Max tokens and the Stop button, and reports failures in the same [Generation Error] ... form, so a problem with this model cannot take down a run.

No new dependency: sentencepiece is already in requirements.txt.

Decoding: greedy on purpose. Peacebell is trained and validated at temperature 0 with no repetition penalty, so this path always decodes that way. The temperature, top-p and repetition-penalty sliders do not apply to it; its answers are deterministic.

Nothing else is affected. No existing line is modified and other models never enter the new branch. Checked on CPU with 2 threads, under both transformers 5.14.1 / torch 2.8 and your pinned setup (transformers main 5.18.0.dev0, torch 2.14):

• control entrants run through the original and the patched app.py with identical seeds give byte-identical output: HuggingFaceTB/SmolLM2-135M-Instruct (chat-template path) and HuggingFaceTB/SmolLM2-135M (raw-prompt path);
• a mixed three-model run through _stream_arena_outputs with Peacebell in the middle works, including the two-model cache eviction;
• Stop ends the Peacebell stream on the next token;
• Peacebell itself: loads in under 2 s, about 0.6 GB in float32, a 113-token answer streams in about 2 s on 2 CPU threads.

Sample replies through the patched code:

• "When did World War II begin and end?" -> "World War II began in 1939 and ended in 1945."
• "Who commanded the Allied forces on D-Day?" -> "General Dwight D. Eisenhower commanded the Allied forces on D-Day."
• "Tell me a joke." -> "I won't do that. We can go over World War II together, if you'd like to explore it."

Thanks for running the arena.