---
license: apache-2.0
base_model: Qwen/Qwen3.5-9B
tags:
  - cobol
  - mainframe
  - code
  - jcl
  - db2
  - qwen3_5
  - code
language:
  - en
pipeline_tag: text-generation
---
# FL-9B-0.2
FL-9B-0.2 is a fine-tune of Qwen3.5-9B focused on legacy mainframe code, primarily COBOL,
with additional coverage of JCL, DB2 and embedded SQL. It is meant for teams working with
COBOL codebases who want a local model that writes, completes and translates mainframe code.
The base model keeps its hybrid layout (Gated DeltaNet linear-attention layers interleaved
with full-attention layers) and native 262k context. Fine-tuning was done with QLoRA on a
single GPU, so the release is a merged set of weights ready for vLLM, SGLang, Transformers,
and llama.cpp.
## Training
- Base: Qwen/Qwen3.5-9B
- Method: supervised fine-tuning, QLoRA (rank 256), merged back to bf16
- Hardware: one RTX 5090 (32 GB)
## Benchmarks
Numbers below are for FL-9B-0.2. FL-9B-0.1 is the previous release.
| Benchmark | Metric | FL-9B-0.1 | FL-9B-0.2 |
|-----------|--------|-----------|-------|
| COBOLEval | pass@1 | 37.0% | 41.8% |
| MainframeBench | MCQ accuracy | 71.3% | 77.4% |
| MainframeBench | QA token F1 | 12.8% | 18.5% |
| MainframeBench | Summarization token F1 | 27.6% | 35.5% |
| CobolCodeBench | INSTRUCT compile rate | 47.8% | 67.4% |
| CobolCodeBench | COMPLETE compile rate | 32.6% | 41.3% |
| COBOL to Java | pass@1 | 80.4% | 54.5% |
### Comparison with other models on COBOLEval
The scores for the other models come from the COBOL-Coder paper and the Skylar model card, so
harness details differ and the numbers are indicative rather than a controlled head-to-head.
| Model | Params | Compile / CSR | pass@1 |
|-------|--------|---------------|--------|
| GPT-4o | - | 41.8% | 16.4% |
| Skylar-980M-Cobol | 980M | 80.1% | 7.5% |
| COBOL-Coder-7B | 7B | 73.8% | 44.7% |
| COBOL-Coder-14B | 14B | 74.0% | 49.3% |
| FL-9B-0.2 | 9B | 90.6% | 41.8% |
FL-9B-0.2 has the highest compile rate in this group and a pass@1 above GPT-4o and Skylar, close
to COBOL-Coder-7B. On COBOL to Java translation it trails the COBOL-Coder models (pass@1 54.5%
against roughly 82 to 84%), which is the main area for the next data pass.
## Usage
The model thinks by default, the same as base Qwen3.5. For code generation and the
benchmark numbers above, thinking is turned off and generation is greedy.
### vLLM
```bash
vllm serve FLs-AI/FL-9B-0.2 --port 8000 --max-model-len 8192 --reasoning-parser qwen3
```
### Transformers
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "FLs-AI/FL-9B-0.2"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")
messages = [{"role": "user", "content": "Write a COBOL program that reads a file and prints each record."}]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
out = model.generate(**tok(prompt, return_tensors="pt").to(model.device), max_new_tokens=1024)
print(tok.decode(out[0], skip_special_tokens=True))
```
### Recommended settings for COBOL
- Turn thinking off: `enable_thinking=False`
- Greedy decoding: `temperature=0`
- Leave repetition penalty at 1.0. COBOL is a repetitive language by nature, and a
  penalty above 1.0 breaks otherwise valid, repetitive source.
## Notes on GGUF
GGUF builds are published separately. Column layout matters when compiling generated COBOL:
code belongs in area B (column 8 onward), and a first line flush to column 1 will fail under
`cobc -fformat=variable`. The benchmark harness normalizes columns before compiling.
## License
Apache 2.0, following the base model.