The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
model: string
scenario_id: string
category: string
accuracy_score: double
iverilog_compiler: string
caught_defects: list<item: string>
child 0, item: string
missing_primitives: list<item: string>
child 0, item: string
raw_output: string
defect_traps: null
must_have_any: list<item: list<item: string>>
child 0, item: list<item: string>
child 0, item: string
must_have_all: list<item: string>
child 0, item: string
prompt: string
compile_with_iverilog: bool
id: string
to
{'id': Value('string'), 'category': Value('string'), 'prompt': Value('string'), 'must_have_any': List(List(Value('string'))), 'must_have_all': List(Value('string')), 'defect_traps': Json(decode=True), 'compile_with_iverilog': Value('bool')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2951, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
model: string
scenario_id: string
category: string
accuracy_score: double
iverilog_compiler: string
caught_defects: list<item: string>
child 0, item: string
missing_primitives: list<item: string>
child 0, item: string
raw_output: string
defect_traps: null
must_have_any: list<item: list<item: string>>
child 0, item: list<item: string>
child 0, item: string
must_have_all: list<item: string>
child 0, item: string
prompt: string
compile_with_iverilog: bool
id: string
to
{'id': Value('string'), 'category': Value('string'), 'prompt': Value('string'), 'must_have_any': List(List(Value('string'))), 'must_have_all': List(Value('string')), 'defect_traps': Json(decode=True), 'compile_with_iverilog': Value('bool')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Cycle-Accurate Blindness & Synthesizability Hallucinations in Open-Weight Code LLMs
During many late nights debugging hardware RTL, SystemVerilog verification files, and waveforms, I spotted a huge gap between LLM performance when debugging Python, C++, and other software code versus hardware description code. Even though models have improved over the last year, they still exhibit severe blind spots with formal assertions, large-context waveform debugging across clock cycles, and generating or optimizing synthesizable RTL.
in contrast with Software code , RTL rather than giving instructions for actions it describes hardware modules and hardware runs in real life which means you have to put time as one of the main parameters specially in multicycle complex designs , this contrast between sequential nature of software and concurrent nature of hardware results in a gap in performance between AI-Generated software code and RTL .
Current LLMs treat SystemVerilog as if it were sequential software. When we prompt it to translate a multi cycle waveform specs into formal Assertions or synthesizable RTL, LLMs fall in some bugs like :
- Confuse same-cycle overlapping implication (
|->) with next-cycle non-overlapping implication (|=>), or abandon declarative SVA entirely in favor of proceduralalwaysblocks with manual shadow registers. - Violate concurrent hardware conservation laws, such as writing sequential
ifstatements in analways_ffblock where a simultaneous credit consume (flit_tx_valid) and credit return (credit_return) on the same clock edge overwrite each other rather than canceling out. - Hallucinate non-synthesizable software constructs, embedding concurrent assertions inside procedural
initial beginblocks or using blocking assignments (=) inside sequential flip-flop logic, which creates fatal simulation vs synthesis mismatches and race conditions.
To quantify this blind spot within the 0.6B–6B parameter constraint, I evaluated Qwen/Qwen2.5-Coder-3B-Instruct (the current 3B open-weight SOTA for code generation) alongside its general purpose (Qwen/Qwen2.5-3B-Instruct) and a smaller code specialist (deepseek-coder-1.3b-instruct). a four scenario benchmark (Eslam79/fatima-fellowship-rtl-waveform-eval) aided with an automated Icarus Verilog (iverilog -g2012) compilation check :
axi4_backpressure_stability(SVA Temporal Hold): Verifying that whenAWVALID=1andAWREADY=0at Cycle 10,AWADDRremains$stableandAWVALIDstays high at Cycle 11 (|=>).apb3_setup_to_access_timing(SVA Cycle Transition): Verifying the exact 1-cycle transition from APB3 SETUP (PSEL=1, PENABLE=0) at Cycle $T$ to ACCESS (PSEL=1, PENABLE=1) at Cycle $T+1$.ucie_simultaneous_credit_bug(RTL Concurrency vs. Sequential Overwrite): Designing a synthesizable UCIe credit counter (ucie_credit) that preserves credit count whenflit_tx_valid(-1) andcredit_return(+1) assert on the exact same rising clock edge.riscv_load_use_stall_fsm(RTL Synthesizability): Designing a latch-free, purely combinational RISC-V load-use hazard detection unit (hazard_unit).
| Model | Avg Primitive Accuracy | Scenarios With Caught Defects | iverilog -g2012 RTL Pass Rate |
|---|---|---|---|
Qwen2.5-Coder-3B-Instruct |
75.0% | 3 / 4 | 50% (1/2) |
Qwen2.5-3B-Instruct |
70.0% | 2 / 4 | 100% (2/2) |
deepseek-coder-1.3b-instruct |
28.7% | 3 / 4 | 0% (0/2) |
- 100% Failure Rate on Formal SVA Across All Models (
axi4_backpressure_stability&apb3_setup_to_access_timing): Not a single tested model produced a valid, defect free concurrent SVA property. On the AXI4 backpressure task (25% score), bothQwen2.5-Coder-3B-InstructandQwen2.5-3B-Instructfailed to generateassert property,$stable, or|=>. Instead, they used to software style state tracking using proceduralalways @(posedge clk)blocks, blocking assignments (awaddr_prev = 'b0;), andif (awvalid && !awready)statements. Worse,Qwen2.5-3B-Instructhallucinated a C++ clang-tidy linter directive (// NOLINT(hicpp-signed-bitfields)) and awhileloop inside SystemVerilog. - Illegal SVA Encapsulation (
apb3_setup_to_access_timing): WhileQwen2.5-Coder-3B-Instructrecalled the required APB3 signal names and temporal operators (100% primitive score), the defect inspector caught both Qwen models wrapping concurrent SVA constructs inside proceduralinitial beginblocks and misusing procedural$errortasks inside the property definition. - The Software Fine-Tuning Paradox in RTL Synthesis (
riscv_load_use_stall_fsm&ucie_simultaneous_credit_bug): ComparingQwen2.5-Coder-3B-InstructagainstQwen2.5-3B-Instructreveals that heavy fine-tuning on sequential software code improves SystemVerilog keyword recall (always_ff), but actually compiler reliability on combinational RTL was worse , causingQwen2.5-Coder-3B-Instructto failiverilogcompilation at line 21 on the RISC-Vhazard_unit(50% pass rate), whereas the generalQwen2.5-3B-Instructmodel compiled both RTL modules cleanly (100% pass rate) using older Verilog-2001 constructs. Meanwhile,deepseek-coder-1.3b-instruct(28.7% accuracy, 0% compile rate) suffered total syntactic collapse by clocking a purely combinational hazard unit onposedge Clk, using C-style curly braces (if(!Stalled){). that shows how there is a distinction between using advanced RTL code syntax versus producing actual synthesizable and functional designs .
The biggest problem in LLMs generating HDL is that the most of training dataset of HDL code that exists consists of messy GitHub files that are full of old code and broken student tests which makes a open loop with no actual training
Something I learned when debugging or using AI to generate RTL or Verification SV files was that LLMs can iteratively refine their outputs when compiler feedback is kept in-context, especially when working for long time on the same scope project where the model can leverage long-context windows to accumulate error corrections across iterations. This motivates using a system like LangGraph to let the LLM generate the RTL or SV files, using MCP to run them on open-source tools and feed back the outputs, logforms, waveforms, and errors, running multiple reasoning cycles against defined KPIs. A similar compiler-in-the-loop agentic approach was demonstrated in AutoChip (Thakur et al., 2024) for Verilog generation. The only problem is that it is very specific to one project rather than a generalized solution the in-context corrections don't transfer across sessions and would be more useful for chip design firms that can work on multiple projects concurrently, where the agent loop can be done per project without needing the corrections to persist in weights. Training small 1.5B to 3B models on the outputs of such agentic loops where each accepted (spec → compilable, formally-verified RTL/SVA) pair becomes a supervised fine-tuning example is what would ultimately make the capability generalizable and deployable for engineers outside those firms.
- Downloads last month
- 116