| --- |
| license: gemma |
| library_name: transformers |
| base_model: |
| - google/gemma-4-E4B-it |
| tags: |
| - gemma |
| - text-generation |
| - instruction-tuned |
| - tool-calling |
| - structured-output |
| - vllm |
| pipeline_tag: text-generation |
| --- |
| |
| # SuperGemma4 E4B Abliterated |
|
|
| `supergemma4-e4b-abliterated` is a private evaluation release whose original |
| upstream base is `google/gemma-4-E4B-it`. |
|
|
| This SuperGemma release is an **abliterated and tuned** derivative of that |
| Google E4B base, with additional work for higher release quality, stronger |
| formatting discipline, better code output, and faster time to first token. |
|
|
| This branch is aimed at users who want: |
|
|
| - strong code and bug-fix behavior |
| - clean JSON and tool-call formatting |
| - fast first-token responsiveness |
| - release-ready serving behavior on Transformers and OpenAI-compatible stacks |
|
|
| ## Why This Build Exists |
|
|
| The original Google checkpoint provides the core Gemma 4 E4B capability base. |
| This project line uses an abliterated release path to reduce refusal-heavy |
| behavior, but that kind of modification can regress on exact formatting, |
| tool-call reliability, and service stability if it is not carefully hardened. |
|
|
| This release focuses on recovering and then surpassing baseline quality where |
| it matters for real usage: |
|
|
| - exact structured outputs |
| - code correctness |
| - bug-fix reliability |
| - server-facing stability |
| - low-friction deployment on Transformers and OpenAI-compatible serving stacks |
|
|
| ## Highlights |
|
|
| - Release-quality score: `92.34` |
| - Exact-eval score: `98.50` |
| - Broad-eval score: `83.10` |
| - JSON exact-match: `100%` |
| - Tool-call accuracy: `90%` |
| - Exact code score: `100%` |
| - Exact bug-fix score: `100%` |
| - Long-context sanity: `100%` |
| - TTFT: `2291 ms` |
| - PREFILL: `2479.70 tok/s` |
| - DECODE: `42.04 tok/s` |
|
|
| ## Lineage |
|
|
| 1. Original upstream base: `google/gemma-4-E4B-it` |
| 2. Abliterated and tuned release: `Jiunsong/supergemma4-e4b-abliterated` |
|
|
| ## Comparison Snapshot |
|
|
| Measured against the same evaluation harness used for: |
|
|
| - `google/gemma-4-E4B-it` |
|
|
| | Model | Release Quality | Exact Overall | JSON | Tool | Code | Bugfix | TTFT ms | PREFILL tok/s | DECODE tok/s | |
| |---|---:|---:|---:|---:|---:|---:|---:|---:|---:| |
| | Google base | 77.46 | 83.50 | 50.0 | 90.0 | 62.5 | 100.0 | 4827.31 | 2456.69 | 42.04 | |
| | SuperGemma4 E4B Abliterated | 92.34 | 98.50 | 100.0 | 90.0 | 100.0 | 100.0 | 2291.23 | 2479.70 | 42.04 | |
|
|
| ## Stability Notes |
|
|
| This candidate was release-hardened against the failure modes that matter in |
| real serving: |
|
|
| - batched OpenAI-compatible serving restored |
| - simple OpenAI-compatible serving restored |
| - unicode output verified |
| - tool-calling output verified |
| - empty-response false-green cases blocked by stricter tests |
|
|
| Validation highlights: |
|
|
| - direct reliability audit: `14/14` |
| - repeat reliability probe: `90/90` |
| - batched soak test: `12/12` |
| - simple soak test: `6/6` |
|
|
| ## Recommended Use Cases |
|
|
| - coding assistant |
| - bug-fix assistant |
| - strict JSON and schema outputs |
| - agent backends that depend on tool-call formatting |
| - standard BF16 deployment on Hugging Face / Transformers stacks |
|
|
| ## Quick Start |
|
|
| ```python |
| from transformers import AutoTokenizer, AutoModelForCausalLM |
| import torch |
| |
| model_id = "Jiunsong/supergemma4-e4b-abliterated" |
| tokenizer = AutoTokenizer.from_pretrained(model_id) |
| model = AutoModelForCausalLM.from_pretrained( |
| model_id, |
| torch_dtype=torch.bfloat16, |
| device_map="auto", |
| ) |
| |
| messages = [ |
| {"role": "user", "content": "Write a compact Python function that groups words by length."} |
| ] |
| |
| inputs = tokenizer.apply_chat_template( |
| messages, |
| tokenize=True, |
| add_generation_prompt=True, |
| return_tensors="pt", |
| ).to(model.device) |
| |
| with torch.no_grad(): |
| outputs = model.generate(inputs, max_new_tokens=256) |
| |
| print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)) |
| ``` |
|
|
| ## Serving |
|
|
| This checkpoint is designed to work well with: |
|
|
| - Transformers |
| - vLLM-style OpenAI-compatible stacks |
|
|
| ## Release Positioning |
|
|
| This private release is the strongest all-around E4B candidate in the current |
| project line for users who want the abliterated base behavior without giving up |
| quality recovery, formatting discipline, or serving readiness. |
|
|
