tracki/companies-dataset

Tracki - Synthetic Companies Dataset

10,196 synthetic companies x 18 columns. Every one of the 15 content fields was written by a language model; nothing is rule-generated. Built for the Tracki final project (RUNI - Intro to Data Science): describe a company, and Tracki returns the 3 most similar companies (embeddings). This would be used for subsribing to their social media and websites.

Every row is fictional. Where the generator's name prior collided with a real trademark it is measured and reported below, not denied.

Schema

Cardinalities and ranges are the observed ones, computed from the published file - not the enum the model was offered. Where the two differ, the gap is itself a finding; see Distribution collapse.

column	source	observed in the data	description
name	LLM	10,196 distinct (100.0% of rows)	Fictional company name (1-3 words). Regenerated in a bulk second pass.
tagline	LLM	10,196 distinct (100.0% of rows)	Marketing tagline (2-8 words). Regenerated in a bulk second pass.
value_prop	LLM	10,196 distinct (100.0% of rows)	One-sentence value proposition. The row's identity and the core of the embedding text.
industry	LLM	15 values, top Fintech 8.2%	Sector. STEERED - asked for explicitly, one per batch.
target_customer	LLM	8 values, top Developers 36.8%	Who they sell to.
business_model	LLM	5 values, top B2B 87.3%	B2B / B2C / B2B2C / D2C / B2G.
revenue_model	LLM	5 values, top Subscription 52.8%	How they charge.
deployment_model	LLM	4 values, top Cloud 97.3%	How the product is delivered.
funding_stage	LLM	8 values, top Series A 13.5%	Ordinal, Bootstrapped -> Public. STEERED - asked for explicitly, one per batch.
primary_market	LLM	3 values, top Global 85.7%	Where they sell.
hq_country	LLM	7 values, top USA 99.0%	Headquarters.
founded_year	LLM	2005 - 2024, 20 distinct	Year founded.
num_employees	LLM	4 - 13000, 61 distinct	Headcount. Band-checked against funding_stage.
funding_total_usd_m	LLM	0 - 3000, 76 distinct	Total raised, USD millions. Band-checked against funding_stage.
annual_revenue_usd_m	LLM	0 - 7500, 121 distinct	Annual revenue, USD millions. Band-checked against funding_stage.
company_age	engineered	2 - 21, 20 distinct	2026 - founded_year. Deterministic - see Redundancy.
stage_rank	engineered	0 - 7, 8 distinct	Integer rank of funding_stage. Deterministic - see Redundancy.
embed_text	engineered	10,196 distinct (100.0% of rows)	Legacy concatenation of 7 columns. NOT the text the recommender embeds - Part 3 measured its categorical prefix (39.5% of the string, only 306 distinct values) as harmful and embeds tagline + value_prop instead. See embeddings/eval_report.md.

How it was built

Three generator notebooks ran in parallel Colab tabs (seeds [42, 43, 44]), each owning 5 of the 15 industries, with Qwen/Qwen2.5-3B-Instruct. Prompts were GPU-batched; every batch was validated and filtered in the loop, and any row breaking a bound was dropped, not repaired.

name and tagline were then regenerated by Qwen/Qwen2.5-0.5B-Instruct in a bulk second pass - 50 items per prompt rather than one, so the model sees and avoids its own previous outputs in context. random picks only which industry and which style to request; it never writes a value.

Two of the fifteen content columns were steered - industry and funding_stage were asked for explicitly, one per batch, because an unsteered pilot collapsed onto Series B. The model still had to write the value, and a mismatch was a rejected row. The other thirteen were left to the model's own choice, which is why they look the way the next section describes.

Deduplication

value_prop is the row's identity - it is what gets embedded and what the recommender compares.

step	before	after	removed
exact value_prop	10,810	10,803	-7
normalized value_prop	10,803	10,782	-21
semantic value_prop (>0.94)	10,782	10,196	-586

The semantic pass (cosine > 0.94 over sentence-transformers/all-MiniLM-L6-v2) removes near-twins that exact matching missed - rows differing only by a trailing period, a swapped preposition, or reordered clauses. They matter because each near-twin would be the other's nearest neighbour by construction.

Row accounting

stage	rows
generation attempts	12,436
rejected by the bounds validator	1,626
written to shards	10,810
surviving the dedupe cascade	10,196
published	10,196

• Row-level validity 86.93% - of everything attempted, the fraction satisfying all bounds. A statement about the validator.
• End-to-end yield 81.99% - of everything attempted, the fraction reaching the published file, after deduplication too. A statement about the pipeline.

Note on earlier releases. A previous version published 2,726 rows: the merge step deduplicated on name as well as value_prop, discarding ~8,000 valid rows whose only fault was sharing a company name with an earlier row, which pinned the dataset size to the generator's name vocabulary rather than to how much data was generated. That filter is gone. A later version shipped this file with a card whose enum sizes described the prompt rather than the data; that is what the present card fixes.

Distribution collapse - what the EDA caught

Thirteen of the fifteen content columns were the model's unsteered choice, and a small instruction-tuned model has a narrow prior. It shows:

column	offered	used	levels never used	most common value
hq_country	15	7	8 never used	USA 99.0%
deployment_model	5	4	1 never used	Cloud 97.3%
business_model	5	5	-	B2B 87.3%
primary_market	7	3	4 never used	Global 85.7%
revenue_model	7	5	2 never used	Subscription 52.8%
target_customer	8	8	-	Developers 36.8%
funding_stage	8	8	-	Series A 13.5%
industry	15	15	-	Fintech 8.2%

Grid-lock

column	distinct values	rows per distinct value	most common	top 10 cover
founded_year	20	509.8	2018 21.3%	95.6%
num_employees	61	167.1	15 11.1%	62.4%
funding_total_usd_m	76	134.2	0.5 11.6%	65.6%
annual_revenue_usd_m	121	84.3	5 9.1%	54.2%

Redundancy

company_age is exactly 2026 - founded_year and stage_rank is exactly the rank of funding_stage - Spearman rho between the first pair is -1.000 by construction. They are kept for convenience; do not feed both members of a pair to a model, and do not read their correlation cell as a result. 8.47% of rows are also exact duplicates across all non-text columns.

Consistency - and whether the validator forced it

The model chose every number. Rows outside the stage band were dropped, so the obvious question is whether a rising ladder could have failed to appear. It could:

funding_stage	funding allowed	funding produced	range used	median	distinct	employees produced	n
Bootstrapped	0 - 1	0 - 1	100%	0.5	9	5 - 20	1,182
Pre-seed	0.05 - 2	0.2 - 1.5	67%	1.2	12	4 - 10	1,362
Seed	0.5 - 8	2 - 6	53%	3.5	9	8 - 35	1,358
Series A	5 - 30	10 - 25	60%	15	9	30 - 100	1,378
Series B	20 - 90	40 - 75	50%	50	8	120 - 350	1,377
Series C	60 - 250	100 - 180	42%	150	11	200 - 750	1,366
Series D+	150 - 900	300 - 750	60%	500	11	400 - 2500	1,303
Public	200 - 5000	250 - 3000	57%	1200	19	1000 - 13000	870

Unaided band compliance, measured over all 12,436 attempts:

check	attempts passing
every band at once	95.05%
funding within stage band	97.79%
headcount within stage band	99.58%
revenue within stage band	99.96%
revenue per head plausible	97.74%
age within stage band	99.99%
stage matches its own numbers	99.99%

Validity - what the model got wrong

Every rejection is a row the model wrote that broke a bound and was dropped, not repaired. All 16 reasons are listed, including the singletons - funding_stage and band_age being near-zero is the point, not a rounding detail.

reason	rows	% of attempts
tagline	331	2.66%
target_customer	301	2.42%
band_rev_per_head	281	2.26%
band_funding	275	2.21%
value_prop	262	2.11%
primary_market	60	0.48%
band_employees	52	0.42%
non_english	22	0.18%
bad_json	14	0.11%
business_model	10	0.08%
revenue_model	9	0.07%
band_revenue	5	0.04%
funding_stage	1	0.01%
deployment_model	1	0.01%
band_age	1	0.01%
hq_country	1	0.01%
total	1,626	13.07%

Known collisions with real company names

The generator's name prior contains real companies. A hand-written blocklist of 168 well-known names finds 10 exact matches (10 rows) and 19 first-word matches (23 rows, 0.23% of the dataset):

Asana, Figma, HubSpot, Instacart, Mixpanel, Mosaic, Salesforce, Slack, Zendesk, Zoom

Those rows are fictional and their figures are fabricated - a row named Slack carries invented headcount and funding and describes an invented product. The blocklist is short and hand-written, so this is a lower bound. Do not treat any name in this dataset as referring to a real organisation. The fix for a future run is to screen generated names against a blocklist at generation time.

What the recommender inherits

• embed_text opens with the sector name in 100.0% of rows, and a row's nearest neighbour shares its industry 98.8% of the time against a 7.1% random baseline. Similarity here is substantially a sector match; compare embedding models on a variant without the boilerplate prefix as well.
• The 0.94 dedupe cut leaves residual near-twins behind:

nearest-neighbour cosine	rows	share
>= 0.95	0	0.00%
>= 0.90	1,336	13.10%
>= 0.85	3,707	36.36%
>= 0.80	5,969	58.54%

Each surviving pair is the other's top-1 recommendation, so a handful of queries will return a near-restatement of the query.

Generation (Part 3+4)

generation/ holds the competitor-feed generator. Given a one-line description of what someone is building, the recommender returns the three most similar companies and the generator predicts the announcement each of them would plausibly publish next, plus the keywords a tracker would match on. This is order to have a free demo of the posts tracking cpabaility (Linkedin, Intsgram, Social Media & News), without calling APIfy's API - a capability that would be included in the subsription version.

• Model: Qwen/Qwen2.5-0.5B-Instruct through HuggingFace transformers. The posts are generated by a HuggingFace model, not an API.
• Announcement types are bounded by funding_stage, the same conditioning idea the dataset itself was built with: a Pre-seed company cannot file for an IPO, a Public one cannot raise a seed round, and a Bootstrapped one has taken no outside money at all. The model still chooses the type, but only from that stage's allowed list, and a choice outside the list is rejected rather than corrected - so the mix of announcement types stays the model's, and an invalid one never ships.
• Tracking keywords are derived, not generated - the company name, the announcement type and the industry are all columns, so nothing in that field can be hallucinated.

Two rounds, not one call

Read this before the validator table, because the shape of the request is what makes the rest work.

The first version asked for all three posts as one nested JSON object. It scored 0 valid feeds out of 60: 70% of failures were format compliance, not judgement. Thirty gate checks in series cap feed validity at 0.95³⁰ ≈ 21%, so the target was unreachable by construction. Tripling the parameter count bought five points; changing the shape bought seventy.

What ships instead is two batched rounds per feed:

round	asks for	output
1	the announcement type, one call per company	~5 tokens, chosen from the stage-filtered list, options shuffled each call so position bias cannot masquerade as the model's prior
2	the post itself, one call per company	prose - no JSON to break

Each round retries up to three times, widening the temperature 0.7 → 0.9 → 1.0 only after a failure. Seven of the first version's reject reasons disappear by construction rather than by relaxing anything.

How the generation prompts work

Everything below lives in tracki_gen_core.py. See the README for why there are two architectures; this file is about what is actually sent to the model.

The system prompts

SYSTEM      (v1)  "…Return ONLY one valid JSON object and nothing else."
SYSTEM_TYPE (v2)  "…Reply with exactly one option from the list, copied verbatim."
SYSTEM_POST (v2)  "…Reply with the post text only - no JSON, no labels, no quotation marks."

Each carries two things and nothing else: the role and the output contract. No company data, no rules, no examples. The system turn is the part that is identical on all ~180 calls, so anything that varies per company is kept out of it.

v2 sends two calls per company

	type call	post call
builder	build_type_prompt	build_post_prompt
output	~5 tokens, one enum word	~75 tokens of prose
system	SYSTEM_TYPE	SYSTEM_POST

The type is chosen first, from that company's stage-filtered list, and then handed to the post call as a fact. Seven of v1's reject reasons disappear by construction: there is no JSON to malform, and type_not_enum / stage_implausible cannot happen.

The user turn carries all the steering

company_line_v2 assembles the per-company block:

ShelfSmart | Retail Tech | Seed | sells to Retailers | B2B / Subscription
   their own tagline (match this voice): "Never run out of your best seller"   ← style sample
   what they do: "Forecasts inventory for independent retailers…"              ← grounding
   posting as: a small team - 'we', concrete about the first real customers    ← STAGE_VOICE
   writing for: Retailers - margins, inventory, storefront operations          ← AUDIENCE_REGISTER
   shape for this post: open with a specific number or milestone               ← POST_SHAPES
   they are announcing: customer_win                                           ← from the type call
   if it states figures: money 0.5-8M (funding_round up to 30M), revenue 0-4M… ← _band_hint

Every line is derived from a dataset column, which is what makes it defensible. The type call gets a shorter version of the same block — no voice, no shape, no band hint — because none of that helps pick an enum and all of it costs tokens.

The rules block is generated from the validator's own constants

f'- {POST_SENTENCES[0]} or {POST_SENTENCES[1]} sentences, {POST_WORDS_V2[0]}-{POST_WORDS_V2[1]} words'
f'- never open with: {", ".join(BANNED_OPENERS[:6])}'
f'- never use: {", ".join(TELL_PHRASES[:8])}'

Change POST_WORDS_V2 and the instruction and the gate move together. Hand-written rules drift: the prompt asked for two sentences and nothing checked it until count_sentences() existed, and threes and fours shipped in the meantime.

Both ban lists are measured, not assumed. TELL_PHRASES are stems ("game-chang"), because banning the whole word shipped the variant. "empower" is deliberately excluded — it appears in 4.6% of the corpus's own taglines, and adding it takes the screen from 6.4% to 10.7% of corpus text, past the point where it stops screening the model and starts fighting the domain.

Few-shot as alternating turns

Shots are emitted as real user/assistant message pairs from FEWSHOT_V2, sampled with a seeded rng — two for the post call, three for the type call. Each shot's user turn is a full company block plus rules, so the model sees the exact shape of the request it is about to answer.

FEWSHOT_V2 is a rewritten pool: v1's examples run 3–4 sentences and would teach the model to fail its own 2–3 sentence validator.

build_type_prompt also shuffles the option list per call. It used to be in identical order every time, and the model used 4 of 13 types across 180 posts with two taking ~96% — position bias and the model's own prior were confounded. Three shots rather than two, from a pool carrying eight different types, for the same reason.

The retry is prompting too

retry_note appends one line naming the miss:

"Your last attempt ran to 64 words, which is too long. Write a new, shorter post of 20-55 words and finish every sentence."

It goes last, after the rules, because it is the one instruction that call exists to obey. The failed draft is discarded, not sent back to be edited — only the fact of the miss travels. The earlier loop re-sent an identical prompt at a higher temperature, which is the wrong lever for a length failure: temperature changes what the model says, not how much.

In one line

The system prompt fixes the role and the output format. Everything else is built from the dataset's own columns and the validator's own constants, so the instructions cannot drift from the gate and every steering choice traces to a column.

Posts are validated, never repaired

The same discipline the dataset was built with. A post that fails any screen is rejected and the reason counted; nothing is patched into shape. A post that still fails after the last retry is shown with its reject reason rather than silently dropped, so a feed can ship partially - two good posts and one honest gap. ok still means all three cleared, so the headline number never flatters the partial case.

screen	rejects a post that
wordcap_post	is outside 20-55 words
n_sentences	is not 2 or 3 sentences
non_english	contains non-ASCII text after quotes and dashes are normalised, so a curly apostrophe is not mistaken for a foreign language
real_company	names a real company
banned_opener	opens with "we're excited", "thrilled to announce" and similar
tell_phrase	carries a machine-written tell - "revolutionary", "seamless", "game-changing" and similar
ungrounded_number	quotes a figure that does not fit a company of that funding stage, checked against the same STAGE_BANDS the dataset was validated with
degenerate_posts	repeats another post in the same feed above cosine 0.90, measured with the same MiniLM encoder that did the retrieval

Type selection has its own two: type_not_enum for a type outside the stage's list, and type_echo for a model that recites the whole list back instead of picking.

Every screen the prompt asks for is a screen the validator enforces. An earlier release did not have that property and published posts carrying hashtags and clichés the prompt had forbidden.

A ninth screen exists and is off: type_mismatch embeds each post against a canonical sentence for its announcement type and rejects the ones that do not read like what they were labelled. It is measured before it is enforced - the same rule Part 1 followed - because MiniLM reads topic as well as announcement and would screen gross mismatch, not classify.

Limitations of the generated posts

• The corpus is synthetic and so are the companies, so nothing here measures whether a human would find a post believable. The validator measures grounding and format, not realism.
• The announcement type is constrained to what a stage allows, so the mix reflects that vocabulary and the model's prior within it - never how often real companies make each kind of announcement. In the notebook the type is drawn uniformly; in the deployed engine the model picks it.
• The real-company screen is a short hand-written blocklist. It is a lower bound, not a guarantee.

Limitations

• Fully synthetic. Every value was written by a small instruction-tuned LLM and validated only against bounds - plausible-looking numbers are not market data and must not be used for any real decision.
• Thirteen of fifteen content columns were unsteered, so their distributions reflect the model's prior, not the real economy. See Distribution collapse for the measured concentrations.
• Rows violating any bound were dropped rather than repaired, so the kept rows are a biased sample of what the model produces. The reject table describes what was thrown away.
• name and tagline were regenerated in a second bulk pass; in the original single pass the model produced only ~2,726 distinct names across 10,803 rows.
• Real company names appear in name - see the collisions section. Rows are fictional regardless.
• company_age and stage_rank are deterministic functions of other columns.
• Shard 3 delivered 2,385 of its 4,200 target, so Edtech / Foodtech / Mobility / HRtech / Gaming carry fewer rows than the other ten sectors.

Files

• companies_final.parquet / .csv - the dataset (SHA-256 991e84089c0bd426...)

• manifest.json - models, seeds, dedupe cascade, counts, fingerprint

• raw/ - the pre-dedupe shards and per-shard reject logs

• embeddings/ - Part 3's sentence embeddings, the recipe, the model comparison and its report

• generation/ - Part 4's competitor-feed generator: config, sample feeds, evidence chart

• notebooks actually present in this repo (8, listed from the repo itself, not by hand):

• generation/tracki_generation_simple.ipynb

• notebooks/companies_fix_and_republish.ipynb

• notebooks/companies_merge_eda_publish.ipynb

• notebooks/companies_p1.ipynb

• notebooks/companies_p2.ipynb

• notebooks/companies_p3.ipynb

• notebooks/companies_rebuild_v2.ipynb

• notebooks/tracki_embeddings.ipynb

EDA