| --- |
| license: apache-2.0 |
| base_model: Tesslate/OmniCoder-9B |
| tags: |
| - code |
| - cobol |
| - assembly |
| - c |
| - legacy |
| - low-level |
| - mainframe |
| - continued-pretraining |
| language: |
| - en |
| --- |
| |
| # Flare-9B v1.1 — Low-Level & Legacy Code |
|
|
| **A 9B coding model that speaks the languages most models forgot.** |
|
|
| Modern code models are great at Python, JavaScript and the usual suspects — and useless |
| the moment you hand them a COBOL payroll routine or a hand-written x86 assembly stub. |
| Flare-9B v1.1 is built for exactly that gap: **COBOL, assembly and C**, the languages that |
| still run banks, mainframes and firmware, with almost no model support behind them. |
|
|
| It's a continued-pretraining pass over [Tesslate/OmniCoder-9B](https://huggingface.co/Tesslate/OmniCoder-9B) |
| on a curated corpus of real low-level and legacy code — tuned to add legacy fluency |
| **without throwing away** the strong general coding ability of the base. |
|
|
| ## Why it exists |
|
|
| Ask a top open coding model to write COBOL and you usually get confident nonsense — missing |
| `DATA DIVISION`, undeclared variables, code that doesn't compile. That's not a knock on those |
| models; there's simply very little clean COBOL in their training data. |
|
|
| Flare-9B v1.1 was trained on a **cleaned, code-only** legacy corpus (no scraped prose, no |
| chat transcripts — just real programs), so it produces COBOL that actually compiles and runs. |
|
|
| ## Benchmarks |
|
|
| All numbers are **pass@1, executed** — every sample is compiled (gcc / g++ / GnuCOBOL / mono) |
| and run against test cases. No self-reported or LLM-judged scores. |
|
|
| | Language | Base (OmniCoder-9B) | **Flare-9B v1.1** | |
| |---|---|---| |
| | **COBOL** | 0% | **50%** | |
| | C | 87% | 81% | |
| | C++ | 87% | 81% | |
| | C# | 75% | 69% | |
| | Assembly x86-64 | 4% | 10% | |
|
|
| The headline: **COBOL goes from completely unusable to genuinely useful** — the base model |
| scores a flat zero, Flare v1.1 solves half of a 40-task executable COBOL suite. General |
| C-family ability stays strong (~80%+), so you get legacy fluency as an addition, not a |
| trade-down. Raw hand-written assembly remains hard for any 9B model, but the simple cases |
| land cleanly. |
|
|
| ## Best use |
|
|
| This model is **completion-oriented** — it shines when you give it code context (a function |
| signature, a program skeleton, a divisions header) rather than long chat instructions. |
|
|
| ```python |
| from transformers import AutoModelForImageTextToText, AutoTokenizer |
| |
| tok = AutoTokenizer.from_pretrained("DarkKnighToS223/Flare-9B-low-level-programming") |
| model = AutoModelForImageTextToText.from_pretrained( |
| "DarkKnighToS223/Flare-9B-low-level-programming", |
| device_map="auto", trust_remote_code=True) |
| |
| prompt = """*> GnuCOBOL: compute 15 + 27 and DISPLAY the result. |
| IDENTIFICATION DIVISION. |
| PROGRAM-ID. MAIN. |
| """ |
| inputs = tok(prompt, return_tensors="pt").to(model.device) |
| print(tok.decode(model.generate(**inputs, max_new_tokens=200)[0], skip_special_tokens=True)) |
| ``` |
|
|
| ## Training |
|
|
| | | | |
| |---|---| |
| | Base | Tesslate/OmniCoder-9B (Qwen3.5, hybrid linear-attention, 9.6B) | |
| | Method | QLoRA (4-bit NF4), LoRA rank 64 | |
| | Data | cleaned code-only corpus: COBOL, x86-64 assembly, C, C++, C# | |
| | Context | up to long-form programs | |
| | Framework | Unsloth + PEFT | |
|
|
| ## Formats |
|
|
| - **GGUF** quants (Q8_0, Q6_K, Q4_K_M, Q2_K) for llama.cpp — use **Q4_K_M or higher** for |
| reliable output; Q2_K on a 9B is best treated as experimental. |
| - Full-precision weights available on request. |
|
|
| ## License |
|
|
| Apache-2.0, inherited from the base model. |
|
|
| --- |
|
|
| *Flare-9B v1.1 — because legacy code still needs to ship.* |
| *im working at way better model than right now...* |
| *need some time...* |
