---
language: en
library_name: unsloth
license: apache-2.0
tags:
  - spec-forge
  - command-runway
  - qwen2.5-coder
  - lora
  - code-generation
  - yaml
base_model: Qwen/Qwen2.5-Coder-7B-Instruct
datasets:
  - githeri/spec-forge-training-data
pipeline_tag: text-generation
---

# qwen2.5-coder-7b-specforge

## Model Description

Fine-tuned [Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) using LoRA adapters on the Spec-Forge training corpus.

The model converts natural-language feature requests into validated YAML specifications that conform to the COMMAND_RUNWAY methodology. Each spec contains:
- `task_id`, `summary`, `depends_on`, `local_goals`, `global_goals_refs`, `context`
- Every `local_goal` has an `Inspect → Create/Modify → Verify` verification flow
- Specs pass a hardened validator (canonical vocabulary, near-duplicate detection, YAML safety)
- Specs are scored against runbook-readiness criteria (hard gate: missing Inspect/Create/Verify stages = 0.0)

## Training Details

| Parameter | Value |
|-----------|-------|
| Base model | `unsloth/Qwen2.5-Coder-7B-Instruct` |
| LoRA rank | 16 |
| LoRA alpha | 32 |
| LoRA dropout | 0.1 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Epochs | 3 |
| Learning rate | 2e-4 |
| Batch size | 1 (effective: 4 via gradient accumulation) |
| Max sequence length | 2048 |
| Quantization | 4-bit NF4 |
| Optimizer | adamw_8bit |
| LR scheduler | cosine |
| Warmup ratio | 0.1 |

## Training Data

- **Source**: 475 seed prompts across 21 feature categories
- **Generation**: Ollama (qwen2.5-coder:7b-instruct) at temperature 0.2
- **Validation**: Hardened YAML spec validator (78 test cases)
- **Scoring**: Runbook scorer with hard gate (0.75 threshold)
- **Format**: Chat format (`system` + `user` + `assistant` turns)

## Evaluation

See `data/eval_results.json` after running `make eval-model`.

Metrics:
- **Validation rate**: percentage of specs that pass the hardened validator
- **Score pass rate**: percentage of specs scoring >= 0.75 on the runbook scorer
- **Target**: >80% score pass rate (held-out prompts)

## Usage

### Ollama (GGUF)

```bash
# Download GGUF from this repo's models/ directory
ollama create specforge -f models/qwen2.5-coder-7b-specforge-gguf/Modelfile
ollama run specforge
```

### HuggingFace Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct", torch_dtype="auto")
model = PeftModel.from_pretrained(base, "githeri/qwen2.5-coder-7b-specforge")
tokenizer = AutoTokenizer.from_pretrained("githeri/qwen2.5-coder-7b-specforge")

messages = [
    {"role": "system", "content": "You are a precise specification generator. Output ONLY a YAML document."},
    {"role": "user", "content": "Add a POST /health endpoint that returns 200 OK"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.2)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Limitations

- Trained on a synthetic corpus generated by the base model itself — quality is bounded by the base model's spec-generation ability
- Specs are scoped to a single-file, single-feature granularity (not multi-stage epics)
- Context is fixed to TypeScript/Express/Prisma/Vitest stack
- GGUF quantization (q4_k_m) introduces minor quality degradation vs the 16-bit merge

## Ethical Considerations

- This model generates structured specifications, not executable code
- All generated specs must pass the hardened validator before use
- Human review is required before feeding specs into a COMMAND_RUNWAY executor

## Citation

```bibtex
@misc{githeri-specforge,
  title={Spec-Forge: From Natural Language to Runbook-Ready YAML Specifications},
  author={Githeri},
  year={2026},
  url={https://github.com/nickrotich/githeri}
}
```

## License

Apache 2.0 — same as the base Qwen2.5-Coder-7B-Instruct model.