No official torrent yet β requesting it tells the flock what to host next. You cannot seed a catalog-only page.
Cataloged from yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1
All models by yuxinlu1 βThis is the full-precision
safetensorsmaster for my Gemma 4 12B coding fine-tune β the same model many of you have been running as GGUF, now in its original weights. π§ π» A focused fine-tune of Gemma 4 12B on verifiable Python coding data: it reasons in the open (edge cases, complexity, approach) and then writes a clean, runnable solution.
This repo holds the un-quantized master weights (model.safetensors, bf16). Use it to:
transformers (needs a recent build with gemma4_unified support).π Just want to run it? You don't need this repo β grab a ready-made quant from the GGUF repo β (runs in ~4.5 GB of VRAM / unified memory in LM Studio, Ollama, llama.cpp, Janβ¦). This master is for builders. π
π v2 is almost here! Initial training of v2 is done and it's in benchmarking + final QA. So many of you flagged the agentic behavior β so this round I significantly grew the dataset (especially agentic data). v2 is focused on agentic + coding. Targeting a release this Friday or Saturday (US Pacific). π
π£ Context length is 256K. This master ships with the corrected max_position_embeddings = 262144 (256K) β the
well-known upstream Gemma 4 metadata bug (config.json once said 131072) is already fixed here, so anything you
quantize/convert from these weights inherits the full 256K. π Thanks to the community member who spotted it!
π§ Thinking mode: it thinks in Gemma's native thought channel before answering (keep
enable_thinking=true, the default chat template handles it). Recommended sampling:temp 1.0, top_p 0.95, top_k 64; for coding you can also go greedy (temp 0) for more deterministic solutions. Needs a recenttransformersthat knows thegemma4_unifiedarchitecture.
All from the GGUF repo:
β οΈ GGUF needs a recent llama.cpp β this is the
gemma4_unifiedarchitecture, older builds won't load it.
There's a tiny Gemma 4 MTP draft model in my main reasoning repo β
MTP/ folder. It's the
stock Gemma 4 drafter, so it pairs with any Gemma 4 12B quant β including these coder quants β for
lossless speculative decoding (byte-for-byte identical output, just faster). Because it's trained on base Gemma 4,
the hit-rate on this fine-tune is a bit lower than on vanilla Gemma 4, but it's free and has no downside. Add three
flags (--model-draft, --spec-type draft-mtp, --n-gpu-layers-draft); see the
main repo for the full command. ποΈ
A distillation of two complementary chain-of-thought sources over verifiable Python coding tasks (algorithmic / function-level problems with deterministic tests):
Real CoT for solid coverage + synthetic "second-attempt" CoT to patch the failures β all verified by execution before training. β
google/gemma-4-12B-it.from transformers import AutoModelForCausalLM, AutoTokenizerimport torch
repo = "yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype=torch.bfloat16, device_map="auto")
msgs = [{"role": "user", "content": "Write a Python function to check if a string is a valid IPv4 address."}]inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)out = model.generate(inputs, max_new_tokens=1024)print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))