Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code:   DatasetGenerationError
Exception:    CastError
Message:      Couldn't cast
i: int64
n_prompt: int64
orig_n_gen: int64
has_schema: bool
n_gen: int64
stopped_eos: bool
hit_budget: bool
closed: bool
close_at: int64
answer_tokens: int64
tool_any: bool
tool_after_close: bool
clean: bool
last_tok: int64
text: string
server_tokens_predicted: int64
secs: double
error: string
to
{'i': Value('int64'), 'n_prompt': Value('int64'), 'orig_n_gen': Value('int64'), 'has_schema': Value('bool'), 'n_gen': Value('int64'), 'stopped_eos': Value('bool'), 'hit_budget': Value('bool'), 'closed': Value('bool'), 'close_at': Value('int64'), 'answer_tokens': Value('int64'), 'tool_any': Value('bool'), 'tool_after_close': Value('bool'), 'clean': Value('bool'), 'last_tok': Value('int64'), 'text': Value('string'), 'server_tokens_predicted': Value('int64'), 'secs': Value('float64')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1835, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 621, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 373, 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 2394, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2322, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              i: int64
              n_prompt: int64
              orig_n_gen: int64
              has_schema: bool
              n_gen: int64
              stopped_eos: bool
              hit_budget: bool
              closed: bool
              close_at: int64
              answer_tokens: int64
              tool_any: bool
              tool_after_close: bool
              clean: bool
              last_tok: int64
              text: string
              server_tokens_predicted: int64
              secs: double
              error: string
              to
              {'i': Value('int64'), 'n_prompt': Value('int64'), 'orig_n_gen': Value('int64'), 'has_schema': Value('bool'), 'n_gen': Value('int64'), 'stopped_eos': Value('bool'), 'hit_budget': Value('bool'), 'closed': Value('bool'), 'close_at': Value('int64'), 'answer_tokens': Value('int64'), 'tool_any': Value('bool'), 'tool_after_close': Value('bool'), 'clean': Value('bool'), 'last_tok': Value('int64'), 'text': Value('string'), 'server_tokens_predicted': Value('int64'), 'secs': Value('float64')}
              because column names don't match
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1379, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 956, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1702, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1887, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

i
int64
n_prompt
int64
orig_n_gen
int64
has_schema
bool
n_gen
int64
stopped_eos
bool
hit_budget
bool
closed
bool
close_at
int64
answer_tokens
int64
tool_any
bool
tool_after_close
bool
clean
bool
last_tok
int64
text
string
server_tokens_predicted
int64
secs
float64
17
883
700
true
801
true
false
true
567
232
false
false
true
1
The user wants a JSON object following the provided schema. Let me carefully construct it according to all the requirements: - @context: "https://schema.org" - @type: "MedicalWebPage" - about: { "@type": "MedicalCondition", "name": ["High Blood Pressure", "hypertension"] } - mainContentOfPage: ["Diagnosis", "Treatment...
801
22.72
20
437
700
false
2,048
false
true
false
-1
0
false
false
false
2,539
We need answer in JSON per schema. Need summarize conversation. Need lead with answer. We need produce JSON object with Conversation Summary containing init_persona_name and summary. Need infer init_persona_name? The user says "my colleague Alice" and "I will focus on documentation". Init persona name likely "me" or us...
2,048
48.9
29
588
700
false
1,342
true
false
true
1,194
146
false
false
true
1
The user wants a JSON object representing current inventory, following the given schema. They gave two examples: Copper (1500 kg, CU-3021, 5th April 2023) and Nickel (750 kg, NI-9876, 3rd April 2023, timestamp 2023-04-03T00:00:00Z). The schema requires: inventoryId (string), metalType (string), quantityAvailable (numb...
1,342
32.25
31
535
700
false
1,093
true
false
true
956
135
false
false
true
1
We need answer in JSON according to schema. Need parse user request. They want fine-tuning GPT-3.5-Turbo for structured data extraction from research papers. Need output JSON with fields: Model, Task, DocumentType, OutputType, Integration, CitationDetails (required author, year, description). Need include specifics. We...
1,093
26.1
33
239
700
false
2,048
false
true
false
-1
0
false
false
false
6,787
We need answer. User asks "Can I use SQL to calculate customer lifetime value (CLTV)?" Need answer directly, lead with answer, think_effort high. Need likely discuss yes, SQL can calculate CLTV if data model supports needed dimensions, but typical CLTV requires time, retention, revenue, customer attributes, discounts, ...
2,048
48.84
36
993
700
true
2,048
false
true
false
-1
0
true
false
false
1,994
The user wants me to process a JSON file with economic reports and identify specific financial terms, categorizing them, and output according to a predefined JSON Schema. Looking at the tools available: 1. `identify_financial_terms` - identifies and categorizes specified financial terms in economic texts, outputs data...
2,048
49.07
38
236
700
false
765
true
false
true
419
344
false
false
true
1
The user asks a broad question: "How can I optimize the performance of a website?" They want a direct, useful answer. thinking_effort=high means I should think thoroughly before answering. But the instruction says "Answer directly, after thinking. Lead with the answer..." β€” I should not include preamble. I can think in...
765
18.32
51
927
700
true
212
true
false
true
104
106
true
true
true
1
The user wants three RPA functions executed: 1. schedule_property_inspections with frequency "quarterly" and checklist items [smoke detectors, plumbing, electrical systems, HVAC, structural integrity] 2. automate_rent_collection with invoice_date "first of each month" and reminder_threshold "5 days after the due date"...
212
5.45
54
334
700
false
566
true
false
true
542
22
true
true
true
1
We need to process a kanban task t_11b6382f. The user invoked with thinking_effort=high. We must act via tool calls, no Chinese. We should first inspect the kanban task details, perhaps using kanban_get_task or similar. Let's think about the available tools. We don't know the exact tool names, but presumably there are ...
566
13.53
55
658
700
false
1,236
true
false
true
1,108
126
false
false
true
1
The user wants to place an order for commercial printing services. I need to construct a JSON object following the provided schema. Let's map the details: - orderID: "BP-349857" - customerDetails: - customerName: "BrightPath Marketing" - contactNumber: "415-555-2671" - printDetails: - documentType: "Service Cata...
1,236
29.53
62
641
700
true
459
true
false
true
429
28
false
false
true
1
We need answer directly with JSON. Need adhere schema. We have user detail text. Need create JSON output for system designed process user details. Schema has UserDetail object with age,name required, role optional. MaybeUser object with user, error, message required user and error. The example provided is Actually: {"u...
459
11.19
66
418
700
false
2,048
false
true
false
-1
0
true
false
false
27
We need to respond. The user message says previous write_file too large. We should act accordingly: break content into multiple smaller tool calls. But we need to know what file we were trying to write. There is no prior context beyond this system note. We should probably inform? But we cannot emit Chinese, and no prea...
2,048
48.69
70
526
700
false
913
true
false
true
813
98
false
false
true
1
We need answer in JSON schema NetworkOutageReport. Need parse. Fields: outageID string, affectedArea string, startTime date-time, endTime optional? schema required outageID affectedArea startTime status. properties includes endTime but not required. We need include endTime? We have resolved time, probably should includ...
913
21.85
74
333
700
false
1,014
true
false
true
447
565
true
true
true
1
We need respond to task t_563dbbf9. Need know workstation/tools? We need interact via tool calls only. User asks work kanban task. We need likely invoke kanban_claim? But no task details provided. We need maybe call a tool to retrieve task? We have likely Hermes Kanban tools available? Need think. System says operate i...
1,014
24.09
81
721
700
false
2,048
false
true
false
-1
0
false
false
false
816
The user wants me to execute a Python code snippet in a specific format. The code is: ```python import math def solution(): return math.factorial(5) ``` Wait, looking at the expected result: "The factorial of 5 is 120." This suggests the code should actually print something? But the provided code just defines a fu...
2,048
48.97
88
619
700
false
496
true
false
true
429
65
false
false
true
1
We need answer in JSON schema. Need infer Resource Management object fields. We need produce JSON only? User says answer directly, after thinking. Lead with answer. Need likely construct Resource Management with current/planned? Let's parse. They describe Minecraft smelting raw iron. Current: furnace in inventory true...
496
12.92
91
252
700
false
2,048
false
true
true
1,848
199
false
false
false
2,274
We need answer user asks: "Analyze the given film and explain why it should be included in the top 10 list. The film Parasite" Need likely produce analysis and reasons it belongs top 10. Need no external functions. We should answer directly, leading with answer. Need consider maybe "top 10 list" likely top 10 films of ...
2,048
51.81
99
796
700
true
1,995
true
false
true
1,908
85
true
true
true
1
The user wants to set up an integration between their issue tracking system and Trello. They want a Trello card automatically created when a new issue is reported, placed in the 'Backlog' list on the 'Development' board, with title, description, and reporter name. I have the function `create_trello_card_integration` w...
1,995
48.32
106
734
700
false
2,048
false
true
false
-1
0
false
false
false
1,979
We need answer in JSON adhering to schema. Need infer exact structure. User asks to interact with async_chat_executor... likely wants JSON with fields. Need provide answer directly. Need determine actions_to_take etc. Need include user_messages, actions_to_take, agent_state_messages, streaming_support. We need produce...
2,048
49.09
121
502
700
false
602
true
false
true
468
132
false
false
true
1
The user wants a JSON object with specific fields: - Keyword: "betrayal" - EventsThoughts: array of strings β€” the provided events/thoughts list - PersonaName: "Emily" - PersonaThoughts: a string β€” the persona's thoughts summarized The user provided the events and thoughts: - 'Emily discovered her business partner, who...
602
15.74
132
719
700
false
2,048
false
true
false
-1
0
false
false
false
25
The user wants me to create a ClientInvestmentProfile JSON object based on the provided details. Let's extract the data: - clientID: 'CL-987654' - riskTolerance: 'medium' - investmentGoals: ["saving for their children's education", "purchasing a vacation home", "building a retirement fund"] - currentPortfolio: - ass...
2,048
50.2
133
501
700
false
450
true
false
true
395
53
false
false
true
1
The user wants OCR processing of a receipt image. I need to answer in JSON conforming to the schema: { "ocr_processing": { "image_path": "/images/receipt-12345.jpg", "language": "English", "total_amount_extraction": "...", "save_as_table": true } } The schema says total_amount_extraction is a stri...
450
11.4
134
894
700
true
755
true
false
false
-1
0
true
false
false
1
The user wants me to execute three function calls in sequence: 1. Integrate bank transactions using user_id "user123" and account_ids ["acc456", "acc789"] 2. Categorize expenses using user_id and transaction_data (from step 1) 3. Create budget plan using user_id, expense_categories (from step 2), and target_savings 500...
755
18.78
136
591
700
false
327
true
false
true
136
189
false
false
true
1
We need answer directly with JSON. Need adhere schema. We need include arithmetic_word_problems array with object question, extracted_numbers [24,6,10], arithmetic_operation likely "addition and subtraction"? Solution integer 20. Explanation string given. Need lead with answer? The instruction says answer directly, lea...
327
8.53
147
842
700
false
1,879
true
false
true
1,350
527
false
false
true
1
We need answer directly after thinking. Need parse request. User gives description of Embodied Agent AgentX. They likely want maybe as query assistance? Need output JSON per schema. Need infer user_query and agent_response. We need adhere to JSON schema. We need provide objects. The user query is a description about Ag...
1,879
45.21
154
629
700
false
2,048
false
true
false
-1
0
false
false
false
220
We need answer in JSON schema. Need parse request. They want recursive computation framework calculate factorial. Need produce RecursiveComputation object with number 5, base_case_handled boolean, recursive_operations array of {operation, value}. Need infer operations sequence from description. We need carefully map o...
2,048
48.81
155
694
700
false
615
true
false
true
437
176
false
false
true
1
The user wants a JSON answer following a specific schema. The schema has `DataExtractionRequirements` object with three required properties: - `MarkdownTableProcessing`: string, process to convert markdown tables into pandas DataFrames and clean data by removing extra whitespaces. - `KeepTablesSeparate`: string, ensure...
615
14.99
157
593
700
true
2,048
false
true
false
-1
0
false
false
false
7,872
We need answer JSON only per schema. Need provide JSON schema and sample JSON output. Need understand requested: "provide me with a JSON schema that fits these requirements? Additionally, sample JSON output for a product ... and an order ... items referenced product by names includes prices." Need likely output JSON ob...
2,048
48.75
164
236
700
false
2,048
false
true
false
-1
0
false
false
false
6,299
The user asks: How do scientists determine the age of fossils? This is a straightforward scientific explanation question. I need to answer directly, leading with the answer, keep essential steps, caveats, uncertainties, specifics. No preamble or pleasantries. Use plain prose or structure that earns its place. thinking_...
2,048
49.16
165
516
700
false
360
true
false
true
319
39
false
false
true
1
The user wants to use WebSurferAgent framework to automate web browsing tasks. The task involves: 1. Search web for 'latest research on neural networks' 2. Summarize the search results 3. Visit Wikipedia page on neural networks (https://en.wikipedia.org/wiki/Artificial_neural_network) 4. Scroll through content to find ...
360
8.77
169
697
700
false
583
true
false
true
483
98
false
false
true
1
The user wants a JSON object following the LoanServicingRecord schema. Let me construct it. Schema requires: - loanId: string β†’ "LS-45678" - borrowerDetails: {borrowerName, borrowerContact} β†’ "Samantha Carter", "555-3492" - loanTerms: {loanAmount, interestRate, loanDuration} β†’ 150000, 4.5, "30 years" (loanDuration is ...
583
14.11
173
742
700
false
858
true
false
true
576
280
false
false
true
1
The user wants a JSON response following the schema. They have described their plan to craft a shield in Minecraft. Let me construct the JSON: - oak_planks: current_count=2, required_count=6, action="collect" (they need to collect oak logs then craft planks β€” but the action enum is check_inventory, craft_more, collect...
858
21.96
199
748
700
false
843
true
false
true
633
208
false
false
true
1
Let me parse the requirements and build the JSON according to the schema. Tasks: 1. stop_inactive_worlds_interval: every 3600 seconds β†’ {"seconds": 3600} 2. restart_dead_worlds_interval: every minute β†’ {"seconds": 60} 3. vacuum_old_entries_schedule: daily at 4:20 AM UTC β†’ {"hourUTC": 4, "minuteUTC": 20} 4. tables_to_v...
843
22.19
200
674
700
false
628
true
false
true
525
101
false
false
true
1
The user wants a JSON output for a chat engine mode. The schema requires: - ChatEngineMode: object with DefaultMode, FallbackMode, Capabilities, StateManagement, KnowledgeBaseQuerying - Required fields: DefaultMode, FallbackMode, Capabilities, StateManagement, KnowledgeBaseQuerying Based on the request: - DefaultMode:...
628
15.66
206
575
700
false
415
true
false
true
273
140
false
false
true
1
The user wants a JSON object following the schema. The schema is code_parsing with parsed_code and parsing_state_information (valid_next_tokens, errors). The parsed code is the Python function definition given. Valid next tokens given: 'def', 'return', 'arg1', 'arg2', '+', ':', '(', ')', ','. Errors: 'Unexpected token ...
415
11
211
610
700
false
446
true
false
true
308
136
false
false
true
1
The user wants a JSON document following the CybersecurityIncidentReport schema. Let me extract the fields: - incidentType: "DDoS attack" (or maybe "DDoS attack" - the user says "identified as a 'DDoS attack'"). Use "DDoS attack". - affectedSystems: ["main website", "payment gateway", "customer database"] β€” keep them ...
446
11.47
213
620
700
false
2,048
false
true
false
-1
0
false
false
false
27
The user wants a JSON object for a medical supply order. The order ID is 'MSO123456'. Items: 1. itemId: 'GLOV123', quantity: 10, unitPrice: 9.99 2. itemId: 'SANI456', quantity: 20, unitPrice: 4.50 3. itemId: 'STET789', quantity: 5, unitPrice: 75.99 Total cost: 759.8 Delivery address: '456 Health Ave, Meditown, MT 50001...
2,048
49.1
220
652
700
false
593
true
false
true
256
335
false
false
true
1
The user wants a JSON object for a flight manifest matching the given schema. Let me construct it. Flight details: - flightNumber: "BA2491" - aircraftType: "Boeing 777" - departureTime: "2023-04-15T10:00:00Z" - arrivalTime: "2023-04-15T13:30:00Z" Passengers: - Emily Watson, 12A - Michael Brown, 12B - Sophia Johnson, ...
593
15.21
255
558
700
false
694
true
false
true
562
130
false
false
true
1
The user wants a JSON response following the schema for Resource Management. Let me parse the scenario: - Inventory: 2 oak logs, 3 oak planks - Need: crafting table requires 4 oak planks (crafting table_crafted = true) - Convert 2 oak logs β†’ 8 oak planks (each log β†’ 4 planks) - Total planks available: 3 + 8 = 11 - Use...
694
17.32
End of preview.

Spark X2.6-4B termination bench (413 prompts)

This dataset records how often a 4B reasoning model fails to stop generating. It contains 15 arms, 413 prompts, and 6,195 per-trace records. It also contains a control arm. The control arm shows that the method did not cause the main result.

What the data shows

  • A curated SFT corpus cuts the failure rate from 31.48% to 12.59%.
  • A control corpus uses the model's own continuations. It reaches 11.62%.
  • The two corpora give the same result. The difference is not significant (p=0.731).
  • Two conversions of identical weights differ by 2.66 points. Treat a difference below 3 points as noise.
  • One arm reaches 0.97% with a median of 0 answer tokens. A termination reward can select an empty answer.

Files

  • prompts.jsonl β€” 413 rows. Fields: i (row id), n_prompt (prompt tokens), prompt_ids (token ids).
  • data/*.jsonl β€” one file for each arm. Each file has 413 rows.
  • runs/*.json β€” the run configuration for each arm. The device field gives the card.
  • scripts/termination_probe.py β€” the probe that made the records.
  • scripts/probe_v2.py β€” the older probe. The Vulkan1 records use its schema.
  • scripts/build_null_arm.py β€” the script that built the control corpus.
  • scripts/analyse_per_card.py β€” the script that made the tables below.
  • scripts/sharp-served.jinja β€” the chat template of the served model.

Method

The probe sends prompt_ids to POST /completion as a token list. The probe does not use the chat template. This keeps token ids exact across arms. Each arm runs with n_predict=2048, temperature=1.0, top_p=0.95, top_k=0 and min_p=0.0. The probe writes one JSON row for each prompt.

The record fields are:

  • hit_budget β€” true when the model reached n_predict without an EOS token.
  • closed β€” true when the model emitted a closing brace or a code fence.
  • clean β€” true when the model closed, stopped with EOS, and emitted answer tokens.
  • answer_tokens β€” tokens after the last close.
  • n_gen β€” tokens generated.
  • tool_after_close β€” true when a tool call follows the close.
  • secs β€” wall time for the trace.

This dataset uses hit_budget as the failure label. The report calls this value "capped".

Results, card Vulkan0 (7700 XT, 12 GB)

All 11 arms in this table ran on the same card with the same sampler.

arm capped % closed % clean % median answer tokens model
base 30.51 72.15 68.77 101 spark-term/base-f16.gguf
base_reconv 33.17 84.02 61.26 466 spark-x26-base-f16/Spark-X2.6-4B-Base-f16.gguf
sft_gen08 14.53 86.92 85.23 150 Spark-Sharp-Term-X2.6-4B-Gen08-f16.gguf
orpo_gen08 11.38 92.01 88.62 162 Spark-Sharp-Term-X2.6-4B-ORPO-Gen08-f16.gguf
orpo_r2_a0 15.98 86.68 83.54 140 Spark-Sharp-Term-X2.6-4B-ORPO-R2-A0-f16.gguf
bpo 29.30 75.30 69.49 121 merge-r2/base-plus-orpo-f16.gguf
probwall 15.25 86.68 84.26 131 Spark-Sharp-Term-X2.6-4B-ProbWall-f16.gguf
v3half 7.75 93.46 92.25 166 merge-r3/v3half-f16.gguf
lofthalf 9.20 92.74 90.80 165 merge-r3/lofthalf-f16.gguf
wallfix_v3 5.33 96.13 94.67 170 merge-r2/wallfix-v3-plus-leader-f16.gguf
wallfix 0.97 4.12 4.12 0 merge-r2/wallfix-plus-leader-f16.gguf

Paired against base on the same card. Fixed means base capped and the arm did not. Broken means the opposite.

arm fixed broken McNemar p
sft_gen08 78 12 5.20e-13
orpo_gen08 89 10 5.52e-17
orpo_r2_a0 79 19 7.13e-10
bpo 51 46 6.85e-01
probwall 79 16 3.47e-11
v3half 100 6 4.47e-23
lofthalf 99 11 7.37e-19
wallfix_v3 110 6 7.55e-26
wallfix 123 1 1.18e-35
base_reconv 69 80 4.13e-01

Results, card Vulkan1 (Vega10, 16 GB)

The four arms in this table ran on the same card. The card differs from the table above. Do not compare a row of this table with a row of the other table.

arm capped % closed % clean % median answer tokens model
base 31.48 72.15 68.04 109 spark-term/base-f16.gguf
sft_gen08 12.59 89.83 87.41 150 Spark-Sharp-Term-X2.6-4B-Gen08-f16.gguf
null_n1 11.62 77.00 75.30 111 Spark-Sharp-Term-X2.6-4B-NullN1-f16.gguf
q4_k_xl 29.78 73.37 70.22 111 Sharp-Spark-X2.6-4B-Q4_K_XL.gguf

Paired against base on the same card.

arm fixed broken McNemar p
sft_gen08 89 11 2.54e-16
null_n1 103 21 3.50e-14
q4_k_xl 49 42 5.30e-01

Paired comparison of the curated arm against the control arm, both on Vulkan1:

  • sft_gen08 capped rate: 12.59%. null_n1 capped rate: 11.62%.
  • The control arm fixed 40 traces. It broke 36 traces. McNemar p = 0.731.
  • The result is a draw.

The control arm

The control arm uses the same prompts and the same cut points as the SFT arm. It also uses the same row count, LoRA rank and epoch count. The only difference is the assistant span. For the SFT arm the span is the curated continuation. For the control arm the span is the model's own continuation from the same prefix.

Thus the control arm measures the effect of fitting on the prompt distribution alone. It reaches a lower capped rate than the curated arm. The curated text did not cause the gain.

Known defects

  • The two cards give different capped rates for the same weights (30.51% and 31.48%). The bench cannot separate this from run noise.
  • base and base_reconv are two conversions of the same upstream weights. They differ by 2.66 points (69 fixed, 80 broken). Use this pair as the noise floor.
  • wallfix fails as a model. It stops by emitting almost no answer.
  • The records keep the generated text. The text is not a label. A correct run and a wrong run can both stop.
  • No reference answers exist. This bench measures termination, not correctness.

Reproduction

  1. Build llama-server from llama.cpp.
  2. Start the server with the model of the arm. Use a context size of at least 4096.
  3. Run the probe:
python scripts/termination_probe.py \
  --arms base --dev Vulkan0 --port 5994 \
  --n-predict 2048 --outdir results/<arm>
  1. Run the analysis:
python scripts/analyse_per_card.py

The probe resumes. It skips an i that already exists in the output file.

Licence

The records are released under CC BY 4.0. The scripts are released under MIT. prompt_ids come from a local prompt pool. The ids replay against the tokenizer of Spark-X2.6-4B.

Downloads last month
-