Spaces:

---

CodeSoft

/

SLM-Arena

Like 20

Running on Zero

Adds wayneworkman2012/peacebell-v1-148M to the main (chat) arena: https://huggingface.co/wayneworkman2012/peacebell-v1-148M

148.55M parameters, trained from scratch, Apache-2.0. It is a single-domain model: it only knows World War II, and it redirects most other topics back to the war. It will lose plenty of general battles and that is fine; it is a first version and I would like it out there.

Why this is a PR and not just a model id

I replayed this Space's own load_models / build_inputs / generate_for_model code against the model locally. It loads fine with your generic loader (AutoTokenizer + AutoModelForCausalLM, trust_remote_code=True, float32), but every reply came back as [Error generating ... 'GPTForCausalLM' object has no attribute 'generate']. Two reasons, both in my repo's design:

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

So it needs a small special case, the same way MetaDiffusion has one.

What this PR changes (app.py: 20 lines added, 0 removed; requirements.txt: 1 line added)

• one entry each in MODEL_IDS, MODEL_DISPLAY, MODEL_PARAMS (148.55e6), GEN_DEFAULTS, MODEL_CONTEXT (16384);
• is_peacebell_model() + generate_peacebell(), and a two-line branch in generate_for_model placed right after the diffusion branch. It honours the same max_new the slider gives every other model;
• sentencepiece>=0.2.0 in requirements.txt (the model's tokenizer imports it).

Decoding: greedy on purpose. Peacebell is trained and validated at temperature 0, with no repetition penalty and no n-gram block, so that is how it is wired (temperature=0.0, no_repeat_ngram_size=0). Its answers are deterministic.

Nothing else is affected. No existing line is modified; other models never enter the new branch. What I checked locally (transformers 5.14.1, your pinned version):

• a control entrant, HuggingFaceTB/SmolLM2-135M-Instruct, run through the original and the patched app.py with the same seeds over 8 prompts: output SHA-256 identical (d45c0901...c02f9f both ways);
• because sentencepiece is a new dependency, I loaded the tokenizer of every model in MODEL_IDS + BASE_MODEL_IDS in a clean environment without and then with sentencepiece installed: tokenizer class, vocabulary size, special tokens, encodings of three test strings and chat-template output are identical for all of them;
• peak GPU memory for Peacebell in float32: 0.59 GiB. The same modeling code already runs on ZeroGPU in my own demo Space.

Unrelated, in case it helps: DedeProGames/Kiyo-135M currently returns 404 on the Hub (it failed to load in both of my runs), so that entry is probably erroring in battles. I did not touch it.

Sample replies through the patched code, 128 new tokens:

• "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.

🤖 Generated with Claude Code