Qwen3-8B — NL2SVA with OL-NL / DFS reasoning
Qwen3-8B fully fine-tuned to translate a natural-language description of a
design property into a SystemVerilog Assertion (SVA), emitting an explicit,
structured derivation inside <think>...</think> before the final assertion.
The reasoning is not free-form chain of thought. Each assistant turn walks a fixed three-part structure derived mechanically from the golden assertion's operator tree:
- OL NL — an operator-level restatement of the requirement, grounded in the design's real signal names and aligned to the assertion's top-level operator structure.
- nl-decomposition tree — a top-down tree, one node per operator, each carrying its natural-language piece and why that operator expresses it.
- operator-merge-sva tree — a bottom-up symbolic derivation labelling every
node
T1, T2, ..., where each line is written in terms of earlier labels.
Prompt format
Trained with the qwen3 chat template. The user turn contains the testbench
module (including a tb_reset wire and a // TODO: ASSERTION marker) followed
by Question: and the property description.
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Jalik/qwen3-8b-codev-sva-ol-dfs-think"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")
messages = [
{"role": "system", "content":
"You are an AI assistant tasked with formal verification of register "
"transfer level (RTL) designs.\nYour job is to translate a description "
"of an assertion to concrete SystemVerilog Assertion (SVA) implementation."},
{"role": "user", "content": """Here is the testbench to perform your translation:
module example(input clk, input rst_n, output valid, output fail);
// ... design body ...
wire tb_reset;
assign tb_reset = (rst_n == 1'b0);
// TODO: ASSERTION
endmodule
Question: Create a SVA assertion that checks: The valid and fail signals must never be active at the same time. Use the signals 'fail' and 'valid'.
You should use `tb_reset` as the disable condition signal. Do not add code to output an error message string.
Enclose your SVA code with ```systemverilog and ```. Only output the code snippet and do NOT output anything else.
Answer:"""},
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True,
enable_thinking=True)
out = model.generate(**tok(text, return_tensors="pt").to(model.device), max_new_tokens=2048)
print(tok.decode(out[0], skip_special_tokens=True))
enable_thinking=True matters: the model was trained with the reasoning block
present and loss computed over it. Disabling it diverges from training.
Output shape
<think>
OL NL: valid and fail must never be true at the same time.
***nl-decomposition tree***
operator: !
natural language piece: valid and fail must never be true at the same time.
reason: The logical NOT (!) negates the conjunction (valid && fail), ...
└── operator: &&
├── operator: null (leaf: valid)
└── operator: null (leaf: fail)
***operator-merge-sva tree***
T1 = fail
T2 = valid
T3 = T2 && T1 [ = valid && fail ]
T4 = !(T3) [ = !(valid && fail) ]
Final assertion (T4): !(valid && fail)
</think>
```systemverilog
asrt_valid_fail_mutex: assert property (@(posedge clk) disable iff (tb_reset)
!(valid && fail)
);
The answer is always the final fenced `systemverilog` block, so downstream
extractors that take the *last* such fence work unchanged.
## Training data
5,000 records sampled from
[`wyt2000/CodeV-SVA-datasets`](https://huggingface.co/datasets/wyt2000/CodeV-SVA-datasets)
(83,195 records), filtered to those whose answer is a single `assert property`
parseable into an operator/signal tree.
The reasoning was generated, not hand-written, and — importantly — **formally
validated**. For each record an OL-NL statement was produced, a *candidate* SVA
was regenerated from that statement alone, and the candidate was checked against
the golden assertion with JasperGold. Only a **`Full equivalence`** result was
accepted; anything weaker (including one-directional implication) was rejected
and the record retried or replaced. So each retained OL-NL statement is known to
be a semantically exact restatement of its assertion, not merely a plausible one.
System and user turns are byte-identical to the source dataset; only the
assistant turn was replaced.
## Training
| | |
|---|---|
| Base | `Qwen/Qwen3-8B` (36 layers, hidden 4096, bf16) |
| Method | Full-parameter SFT (not LoRA) |
| Framework | LLaMA-Factory 0.9.3.dev0, DeepSpeed ZeRO-3 |
| Hardware | 4 × NVIDIA H200 |
| Epochs | 3 (1,875 steps) |
| LR | 1e-5, cosine, warmup ratio 0.1 |
| Effective batch | 8 (1 × 2 grad-accum × 4 GPUs) |
| Max sequence | 16,384 tokens |
| Precision | bfloat16 |
| Final train loss | **0.1725** |
| Runtime | 68 min |
Loss fell from ~1.76 to ~0.12 with gradient norm settling from ~35 to ~0.5.
## Limitations
- **Convention-bound.** Training data consistently uses `tb_reset` as the
`disable iff` condition and a testbench with a `// TODO: ASSERTION` marker.
Prompts departing from that shape may degrade.
- **Single-assertion scope.** Trained only on examples whose answer is exactly
one `assert property`. Multi-assertion or multi-clock requests are out of
distribution.
- **Not verified end-to-end here.** Training loss is not correctness. The
reasoning traces in the *training data* were JasperGold-verified, but this
model card reports no benchmark score for the fine-tuned model itself — treat
generated assertions as proposals to be formally checked, not as correct by
construction.
- **Reasoning is generated.** The decomposition trees come from an LLM pipeline;
the operator structure is mechanical, but the natural-language justifications
in the `reason:` fields were not individually reviewed.
## License
Released under Apache-2.0, matching the `Qwen/Qwen3-8B` base model. Training data
derives from `wyt2000/CodeV-SVA-datasets`; consult that dataset for its own terms.
- Downloads last month
- -