coppead/causalista-4b

Causalista 4B

A small, local model for one decision in causal analysis. Give it a table's column dictionary, say which column is the intervention and which is the outcome, and tell it how the data were generated. It answers which columns to adjust for to measure the average effect, or says that the effect is not identifiable with these columns and names what is missing. It never sees the rows: a separate runner, included in the code repository, estimates the effect with the columns the model chose.

Em português: um modelo pequeno, que roda na máquina da pessoa, treinado para uma decisão só: diante do dicionário de uma tabela e da história de como os dados foram gerados, dizer por quais colunas ajustar para medir o efeito de uma intervenção, ou dizer que não dá e o que falta. A conta é feita por um programa separado. Código, dados e páginas: github.com/coppead/causalista e coppead/causalista-worlds.

• Developed by: André Fonseca, COPPEAD Graduate School of Business, Federal University of Rio de Janeiro (UFRJ), with an AI coding assistant
• Model type: Qwen3-4B (decoder-only, 4B parameters) with two LoRA adapters merged in; answers in JSON
• Languages: Portuguese and English (trained); Spanish (never trained, reported as a transfer test)
• License: Apache 2.0
• Fine-tuned from: Qwen/Qwen3-4B, revision 1cfa9a7208912126459214e8b04321603b3df60c; tokenizer unchanged
• Version: 0.1.0, released 2026-10-25 (code frozen and files uploaded on 2026-10-24); versions below
• Contact: https://github.com/coppead/causalista/issues
• Code, data, quantized files: github.com/coppead/causalista · coppead/causalista-worlds · coppead/causalista-4b-GGUF

At a glance (final test, 1,515 held-out worlds in three languages; details under Evaluation): this model 99% accuracy (score +0.98); the base without fine-tuning 47% (−0.05); a 27B open model with reasoning 99% (+0.99, 152-world sample). On 48 cases that imitate human prose: 69% for this model, 96% for the 27B. Every recorded answer is in the dataset under runs/.

It is the 0.1 release of the Causalista project at COPPEAD/UFRJ: Qwen3-4B, fine-tuned with supervision on 22,906 synthetic worlds with a known answer, then given a short reinforcement step (GRPO, group relative policy optimization) whose only reward was the verifier's score on cases that simulate text written by people. Everything about it is reproducible from the code repository (github.com/coppead/causalista) and the dataset (coppead/causalista-worlds); quantized GGUF files are in coppead/causalista-4b-GGUF.

The task

The model sees: the question (the average effect of column T on column Y), a story of how the data were generated (who received the intervention and why, what was measured before and after it, what is known not to be in the table), and the column dictionary with a simple summary of each column. Instructions are in English; story and dictionary are in the user's language (trained in Portuguese and English; Spanish was never trained and is reported as a transfer test).

The model answers, in JSON: the role of each column (confounder, mediator, collider or neutral) with a one-line reason; the verdict (identifiable by adjustment, or not_identifiable with these columns); the adjustment set, or what is missing; and the assumptions it relies on.

The runner then estimates the effect with the chosen columns, using a doubly robust estimator with cross-fitting (DoubleML's interactive regression model), and reports the effect, its interval, the naive difference for contrast, and the assumptions. When the verdict is "not identifiable", the report says why and what data would resolve it.

The prompt builder, the answer parser and the scorer are in src/causalista/task/ of the code repository. An answer is scored by a verifier that knows the true causal graph: +1 when the adjustment set is valid (or when the abstention is right), −1 for a confident error (asserting an effect with a biased set, or when the effect is not identifiable), −0.5 for an undue abstention, −1 for an answer the runner cannot use.

Intended use

Direct use. Deciding which columns of a table to adjust for to measure the average effect of a binary intervention on an outcome, or saying that the effect is not identifiable with the columns at hand, in tables of the kind management theory describes: customer, employee, student or client records with an intervention and an outcome. The decision feeds a separate estimator; the model never computes the effect itself.

Primary users. Analysts and researchers who already hold a table and a story of how it was generated, and want a first, fast, local opinion on what to adjust for, with the assumptions written out. The repository's scripts/ask.py is the intended entry point.

Out-of-scope use. It is not a general assistant and was not measured on anything but this task. It should not decide alone on credit, hiring, health coverage or any decision about a person: its output is one input to a human analysis, and the estimator's assumptions must be read. Outside the eight management themes and the prompt format it was trained on, nothing here says how often it is right. It does not recover effects of real experiments: every number below is on synthetic data with a known answer.

Training

Base	Qwen/Qwen3-4B, revision 1cfa9a72
Method	supervised fine-tuning with LoRA (low-rank adapters; rank 16, alpha 16, all linear projections), loss on the answer only; one run, seed 0; no hyperparameter search (the training script's defaults)
Examples	22,906: 19,814 stories told one fact per sentence (9,719 pt, 10,095 en) and 3,092 in prose written by open models (1,943 pt, 1,149 en), from six management themes (telecom churn, retail, bank credit, subscription software, HR, education)
Targets	the ideal answer derived from the verifier's answer key; no answer was written by a language model
Schedule	1 epoch, 1,432 steps, batch 16, learning rate 2e-4 linear with 5% warm-up, bf16, Unsloth + TRL
Hardware	one RTX 4090 (24 GB) under WSL2, 2 h 22 min
Reinforcement step	GRPO (Shao et al., 2024) with the verifier's score as the only reward (Dr. GRPO loss of Liu et al., 2025, no KL term): 8 answers per case, 2 cases per step, learning rate 2e-5; stopped at step 20 (14 minutes) by the guard against collapse into always abstaining
Reinforcement data	208 cases that simulate text written by people (128 pt, 80 en), cases/treino-simulados/ in the dataset; 32 more cases were held out as a development test
Published weights	the base with both adapters merged (safetensors); the two adapters under adapter/sft and adapter/grpo-c; the training records and the record of the model choice under training/

Two themes, e-mail marketing and private health insurance, were held out of training entirely, as were the 888 graph structures of the two test sets. The exact training file is examples/sft-final.jsonl in the dataset.

Reinforcement learning. The plan was supervised training followed by reinforcement with the verifier's score as the only reward (GRPO). On 2026-10-07, three runs on synthetic worlds produced no measurable change: from the supervised model, 200 steps changed nothing measurable (it already answered the synthetic worlds almost perfectly); from the untrained base, with and without a written reasoning step, the runs were stopped by the guard against collapsing into "not identifiable" after 35 and 28 steps. Reinforcement can only teach where the model still errs, and that is text that imitates people (the supervised model scored 67% there). So on 2026-10-18 two arms were trained on 208 such cases, starting from the supervised model: C, reinforcement with the verifier's score (this model), and D, supervised training on the same 208 cases with their ideal answers. The choice between the supervised model, C and D followed a rule written before the results: the highest score on 32 held-out cases, provided the arm loses no more than 2 points of accuracy on the synthetic development set. D reached 94% on the 32 cases but lost 5 points on the synthetic worlds and was excluded by the rule; C reached 75% against 72% for the supervised model, with no loss, and was chosen. In short: the reinforcement step changed one case in 32 and, on the final measurement below, one case in 48, and it was stopped early by the guard. All runs are recorded in the repository.

Compute and environmental impact

One RTX 4090 for the two published runs: 2 h 22 min of supervised training and 14 min of reinforcement, about 1.2 kWh at the card's rated power as an upper bound, not counting the exploratory runs recorded in the repository. Emissions depend on the local grid and were not estimated.

Evaluation

All numbers come from the harness in the repository, with the same prompt and the same estimator for every responder; every answer is recorded in the dataset under runs/, and the exact commands to regenerate each row are in the repository README ("Reproducing the results"). One training run and greedy decoding, so the evaluation is deterministic; the uncertainty that remains is sampling: at 99% accuracy on 1,515 worlds the binomial 95% interval is about ±0.5 points, on 48 cases about ±13 points. Results are broken down by language, by theme (seen or held out), by identifiability and by kind of story. A frontier model was not measured.

Role	Model	How it was run
Base and fine-tune	Qwen/Qwen3-4B at 1cfa9a72	transformers 5.5, bf16, greedy, thinking off
Naive strategies	"adjust for everything" (what estimation libraries do when given all columns) and "adjust for nothing" (the naive difference)	rule:all, rule:none in the harness
Larger comparator	Qwen3.8 27B, qwen3.8:27b through Ollama	thinking on (its default), 1 world in 10
Prose writers and readers, translators	Qwen3 14B, Gemma 4 26B, Granite 4.1 30B, Qwen3.5 9B and 35B, through Ollama	named per set in the dataset card

Set names: placar-v3 is the development set ("placar" = scoreboard; v3 = third generation of the world factory), teste-v3 the final test set, opened once. Column names are Portuguese identifiers in all three languages, and the verifier's reasons in the records are in Portuguese; docs/README.md in the repository has the glossary.

The measurement protocol was written before the results (section 31 of docs/status.md in the repository, amended on 2026-10-07 for the arms trained on cases). The development set placar-v3 and the 32 held-out cases were looked at freely. The final test teste-v3 was opened once, on 2026-10-18, for the table below. The 48 cases that simulate text written by people were frozen before any model answered them and never used for tuning; the supervised model, the base and the comparators had been measured on them once during development, in breach of the protocol, and this model was measured on them only after being chosen by the rule.

Final test: 505 worlds in each language, eight themes, one fact per sentence

Two of the eight themes never appeared in training; Spanish never appeared in training.

Responder	Score	Accuracy, 1,515 worlds	Portuguese	English	Spanish (never trained)	Two themes never trained (381)	Not-identifiable worlds
Adjust for everything	−0.66	17%	17%	17%	17%	18%	0%
Adjust for nothing	−0.95	2%	2%	2%	2%	2%	0%
Base without training (Qwen3-4B)	−0.05	47%	46%	46%	47%	47%	4%
Supervised model only	+0.98	99%	100%	98%	98%	96%	100%
This model (supervised + reinforcement step)	+0.98	99%	100%	99%	98%	96%	100%
Qwen3.8 27B with reasoning, 1 world in 10 (152)	+0.99	99%	100%	100%	98%	100% (37)	100%

This model made two undue abstentions (0.1%) and no invalid answer; its other 14 errors are confident errors in identifiable worlds. All 16 errors are in English and Spanish.

Development set: 384 worlds in each language, plus prose

Accuracy of this model and, in parentheses, of the base without training:

Story	Portuguese	English	Spanish (never trained)
One fact per sentence, 384 worlds	100% (47%)	98% (47%)	98% (46%)
Prose written by a model used only for the test sets (Granite 4.1 30B)	96% (39%), 113 worlds	98% (35%), 116 worlds	not available

Mean score over the 1,381 worlds: +0.97 for the supervised model, −0.09 for the base; on the 384 Portuguese one-fact worlds, −0.77 for "adjust for everything" and −0.93 for "adjust for nothing". This model, after the reinforcement step: 98% over the 1,381 worlds, 100% in Portuguese one fact per sentence, 96% in Portuguese prose. The 27B open model with reasoning answered all 153 worlds of its sample correctly (96 one fact per sentence, 57 prose).

32 held-out cases that simulate text written by people (development)

Responder	Score	Accuracy
Supervised model only	+0.48	72%
This model (arm C)	+0.56	75%
Arm D, supervised on the 208 cases	+0.91	94%, but 95% instead of 100% on the synthetic worlds: excluded by the rule

48 cases that simulate text written by people

E-mails, chat messages, meeting notes and report excerpts, written by a large language model in the role of managers and analysts, on themes inside and outside the eight; 24 in Portuguese, 16 in English, 8 in Spanish. They are not text written by people.

Responder	Score	Accuracy	pt	en	es
Adjust for everything	−1.00	0%	0%	0%	0%
Adjust for nothing	−0.75	12%	12%	12%	12%
Base without training	+0.05	52%	50%	56%	50%
Supervised model only	+0.36	67%	75%	56%	62%
This model (supervised + reinforcement step)	+0.41	69%	75%	62%	62%
Qwen3.8 27B with reasoning	+0.92	96%	100%	94%	88%

The supervised model's 16 errors: 8 times it left out a confounder that was stated indirectly; 5 times it asserted an effect in a case that was not identifiable (none in Portuguese); 3 times it abstained without need. The reinforcement step fixed one English case.

What the numbers say

In synthetic worlds, including two themes and a language it never saw, this 4B model matches a 27B open model and is far above the base and the naive strategies. In text that imitates people, it is 27 points below the 27B model: reading what a person writes is where it still fails, and that is where the project's work now lies. The reinforcement step, the project's thesis, changed one case in 32 and one in 48; the supervised arm trained on the same cases learned far more about that text (94% on the 32 cases) at the cost of 5 points on the synthetic worlds, which suggests that the next step is supervised training that mixes both kinds of story, not more reinforcement on this signal.

Bias, risks and limitations

The model gives a confident, structured answer to any table that fits the prompt format, including tables far from the eight management themes it was trained on; outside that family, nothing here says how often it is right. Its answer is a decision a person may act on, so the runner reports the assumptions with every estimate and the record of every measured answer is public. The stories it was trained on were written by language models from theory sheets, and the "human" text it was measured on is simulated; both carry the biases of the models that wrote them.

What is not claimed:

• That the model beats a frontier model: no frontier model was measured.
• That it reads real text written by people: the cases are simulated by a language model.
• That it recovers effects of real experiments: every number above is on synthetic data with a known answer.
• That the task is hard for large models: a 27B open model answers nearly every synthetic world. The result is about what a small local model can learn from a verifiable signal, not about the state of the art.
• That the estimator's interval is calibrated outside this family of worlds: its 95.3% coverage was measured on the project's own synthetic worlds, not on other data-generating processes.

How to use

The model expects the exact prompt format it was trained on. For a table of your own, scripts/ask.py in the code repository builds the task from a CSV, a story of how the data were generated and a one-sentence dictionary of the columns, asks the model and runs the estimator on its answer (ollama pull hf.co/coppead/causalista-4b-GGUF first; no GPU needed). Through Ollama, the harness renders the prompt itself with thinking off and sends it raw, because Ollama's own think: false does not close the thinking block for this model (see the GGUF card):

uv run --no-sync python scripts/ask.py --csv table.csv --treatment email --outcome bought \
    --story story.txt --dictionary columns.yaml --model ollama:hf.co/coppead/causalista-4b-GGUF:nothink

To measure the model on the saved worlds, the harness builds the prompt from a saved world or a written case and runs the model locally:

uv sync
hf download coppead/causalista-worlds --repo-type dataset --include "worlds/placar-v3/*" --local-dir outputs
uv run python scripts/run_models.py --worlds placar-v3 --every 20 --model hf:coppead/causalista-4b:nothink

With transformers directly, build the prompt with the package (here from the first written case; make_task(...) builds one from your own table) and decode greedily, with thinking turned off. Requires transformers 5.0 or later: the files were saved with 5.5; with 4.x the tokenizer does not load and config.json is read with the wrong rope_theta, which degrades the output silently.

from pathlib import Path
from transformers import AutoModelForCausalLM, AutoTokenizer
from causalista.harness import case_worlds                 # pip install git+https://github.com/coppead/causalista
from causalista.task import build_prompt, parse_answer

task = case_worlds([Path("cases/fase0.yaml")])[0].task     # the first written case of the repository
tok = AutoTokenizer.from_pretrained("coppead/causalista-4b")
model = AutoModelForCausalLM.from_pretrained("coppead/causalista-4b", dtype="bfloat16", device_map="auto")
prompt = build_prompt(task)
text = tok.apply_chat_template([{"role": "user", "content": prompt}], tokenize=False,
                               add_generation_prompt=True, enable_thinking=False)
out = model.generate(**tok(text, return_tensors="pt").to(model.device), max_new_tokens=2048, do_sample=False)
answer = parse_answer(tok.decode(out[0], skip_special_tokens=True).split(prompt)[-1], task)

The rows of the table never go to the model. To get the effect, give the chosen columns to the runner (causalista.estimator).

The adapters. adapter/sft is a LoRA adapter on Qwen/Qwen3-4B. adapter/grpo-c is a LoRA adapter on the supervised model, that is, on the base with adapter/sft already merged; it is not an adapter on the bare base. Its adapter_config.json names the local folder where that merged model lived during training, so load it by passing the model object to PEFT instead of relying on that path:

from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B", dtype="bfloat16", device_map="auto")
supervised = PeftModel.from_pretrained(base, "coppead/causalista-4b", subfolder="adapter/sft").merge_and_unload()
published = PeftModel.from_pretrained(supervised, "coppead/causalista-4b", subfolder="adapter/grpo-c")

The merged model.safetensors at the root of this repository is exactly that result, so most users need neither adapter.

Provenance

• Answers used in training always come from the verifier, never from a language model.
• The sentences of the training stories come from eight theory sheets written by a large language model from management theory; one was approved by the author, seven were reviewed sentence by sentence by independent instances of the same model and approved on the author's behalf. Prose was written by open models (Qwen3 14B, Gemma 4 26B) and kept only when two other open models agreed it was faithful.
• The simulated human-like cases, both the 208 used in the reinforcement step and the 48 used for measurement, were written and reviewed by instances of a large language model.
• The author made the design decisions: the family of worlds and its ranges, the themes, the score weights, the base model, the training size and the measurement protocol.

Full history, decisions and every measurement: docs/status.md in the repository (in Portuguese).

Versions

Version	Date	What
0.1.0	2026-10-25 (files uploaded 2026-10-24)	first release: supervised fine-tuning plus the reinforcement step described above

The code repository has its own patch versions (0.1.1 on 2026-10-25: scripts/ask.py, the Ollama thinking-off fix, documentation after the audits); they do not change the weights. New model versions get a new tag in the code repository, a new entry in its CHANGELOG.md, and a new revision here; errata are listed in the CHANGELOG.

License and citation

Apache 2.0, like the base model.

@software{fonseca2026causalista,
  author    = {Fonseca, André},
  title     = {Causalista: a small, local model that decides what to adjust for and says when the data cannot answer},
  year      = {2026},
  version   = {0.1.0},
  publisher = {COPPEAD/UFRJ},
  url       = {https://github.com/coppead/causalista},
  note      = {Model: \url{https://huggingface.co/coppead/causalista-4b}}
}

Questions and problems: open an issue at https://github.com/coppead/causalista/issues.