--- license: apache-2.0 language: - en base_model: javierbits/Javi0.3-4B base_model_relation: finetune pipeline_tag: image-text-to-text library_name: transformers tags: - qwen3.5 - tool-use - agent - weight-space - dgx-spark - javi --- ![Javi0.4-4B](https://huggingface.co/javierbits/Javi0.4-4B/resolve/main/javi04-cover.png) # Javi0.4-4B (control) Javi0.4-4B is **Javi0.3-4B with the same embedding row changed**: token 51187 (`=open_page`) is multiplied by 0.84. No fine-tuning, no new data, no gradients. Every other weight is byte-identical to Javi0.3-4B. It exists as the 4B control for the Javi0.4 weight-space experiment: the same row edit applied to the larger model. ## Measured effect On 8 held-out eval-v3 canonical `fail` questions with the first `web_search` forced to fail, Javi0.3-4B already chose `web_search` in 4/4 cases, so that probe sees no change on 4B. On an opencode 403 replay, both models fetch another typed URL instead of searching again. ### Full eval v3 175 held-out seeds, 3 runs, date pin 2026-10-07, cached search results. Javi0.3-4B answers on the same seeds are the baseline. Judges: DeepSeek-V4.1-Flash-1017 and mimo-v26-dual. | metric | Javi0.3-4B | Javi0.4-4B | |---|---:|---:| | core pass (judge average) | 0.50 [0.45, 0.55] | 0.56 [0.52, 0.61] | | core pass (deepseek) | 0.64 | 0.74 | | core pass (mimo) | 0.36 | 0.39 | | recovery after a tool failure | 0.58 (n=12) | 0.75 (n=12) | | made-up links | 0.01 | 0.00 | | loops | 0.03 | 0.02 | Paired core pass difference: **+0.06 [0.03, 0.10]** (deepseek +0.10 [0.05, 0.14], mimo +0.04 [0.00, 0.07]), 95% bootstrap over 175 seeds. The deepseek interval clears zero; the mimo interval touches it. Conflict questions are the one category down (0.26 vs 0.40, n=7, directional at that n); unknowable questions are up (0.55 vs 0.32, n=10). Two answer episodes (2 seeds, one per run) carry no deepseek verdict because the judge omits the `tools` or `style` key; both seeds are judged in the other runs, so the table covers 175 seeds. Incomplete episodes (7 across the three runs) are scored 0 for core pass and skipped by the judges. Full scoreboard: `data/v2/eval/javi04-vs-03/score.md` in the source tree. This control measures slightly above Javi0.3-4B, but the targeted probe cannot attribute the gain to the `open_page` row alone. The 4B's remaining recovery gap in the `webfetch`/`web_fetch` route is untouched by the edit. ## Files Merged bf16 safetensors (two shards, 8.8 GB), tokenizer, processor files and chat template. The vision tower is Qwen3.5-4B's, unchanged; the edit is in the language model only. ## Run it Thinking off, temperature 0.2, top-p 0.8, top-k 20. **GGUF** for llama.cpp or lmdeck: [Javi0.4-4B-GGUF](https://huggingface.co/javierbits/Javi0.4-4B-GGUF). **lmdeck** ```bash lmdeck use javi0.4-4b ``` **Transformers** ```python from transformers import AutoModelForImageTextToText, AutoProcessor model_id = "javierbits/Javi0.4-4B" processor = AutoProcessor.from_pretrained(model_id) model = AutoModelForImageTextToText.from_pretrained(model_id, dtype="auto", device_map="auto") messages = [{"role": "user", "content": [{"type": "text", "text": "What's the latest stable Rust release?"}]}] inputs = processor.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, tokenize=True, return_dict=True, return_tensors="pt").to(model.device) out = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.2, top_p=0.8, top_k=20) print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)) ``` ## The edit ```python embed_tokens[51187] *= 0.84 # =open_page ``` Token 51187 is a single token in the Qwen3.5 tokenizer. The row is tied between input and output embeddings, so the output logit for `=open_page` is reduced. ## Links - Base and recommended 4B: [Javi0.3-4B](https://huggingface.co/javierbits/Javi0.3-4B) - 2B experiment: [Javi0.4-2B](https://huggingface.co/javierbits/Javi0.4-2B) - Apache 2.0, inherited from [Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B).