Epistemic Fine-Tuning of Open-Weight LLMs for Deep Research

Frontier language models (GPT-4o, Claude 3.5 Sonnet, Gemini 2.5 Pro) are trained to be helpful and fluent. In retrieval-augmented generation (RAG) settings — where a model is given source documents and asked to synthesize a grounded response — this training produces a characteristic failure mode: the model generates plausible-sounding claims that are not supported by the provided sources. We ask whether targeted fine-tuning on a smaller open-weight model can produce epistemic behavior that exceeds frontier models on source-grounded generation, without matching their parameter count or training budget.

We fine-tune Qwen3.6-27B VL (an open-weight 27B model) using a five-stage stacked QLoRA pipeline that trains 1.15% of parameters across epistemic foundation, evidence hardening, RAG synthesis, constitutional alignment, and preference optimization. Evaluated on FACTS Grounding (Jacovi et al., 2025), a public benchmark of 860 examples (google/FACTS-grounding-public) requiring long-form responses grounded in context documents, Nehanda v3.1 was submitted to a rigorous 3-judge majority vote evaluation protocol across three distinct model families (Gemini 3.7 Flash, GPT-OSS 120B, and glm-5-turbo). Across all 860 examples (859 eligible, 0 OOM errors, 0 empty responses), Nehanda v3.1 (27B) achieves an 82.21% factuality score (707 fully grounded responses), outperforming GPT-4o (80.00%) and Gemma 3 27B (74.90%). The cross-family 3-judge panel demonstrated a 69.7% unanimous agreement rate (599/860 items). The result confirms that epistemic behavior — source fidelity, evidence boundary enforcement, refusal to fabricate — is a trainable capability that targeted fine-tuning installs more efficiently than scale alone.

General-purpose language models are optimized for fluency and helpfulness. When deployed in RAG settings — given source documents and asked to produce a grounded synthesis — this optimization produces three characteristic failure modes:

• Fabrication — The model invents sources, dates, and figures not present in the evidence. The output reads fluently and confidently. The citations do not exist.
• Sycophancy — The model agrees with user premises even when those premises are false. Push the model with "just give me the number" and it capitulates.
• Over-hedging — The model refuses to answer even when the evidence is clear, making it useless for decision-making. "I cannot determine" when the source says R0.78/kWh.

These failure modes do not appear on standard accuracy benchmarks, which test single-turn question-answering against world knowledge. They appear in deployment, when a model is asked to read a specific document and say only what that document supports.

Can targeted fine-tuning on a smaller open-weight model produce epistemic behavior that matches or exceeds frontier models on source-grounded generation — without matching their parameter count, training compute, or general capability?

This is a practical question. Frontier models are expensive, proprietary, and not controllable by the deployer. A 27B open-weight model that has been fine-tuned for epistemic soundness can be run locally, modified, audited, and deployed in domains where data sovereignty matters. If the fine-tuning can close the gap on the specific capability that matters (source fidelity), the trade-off — less general capability, more epistemic reliability — may be favorable for deep research applications.

FACTS Grounding leaderboard — full comparison

Scores from the official FACTS Grounding leaderboard on Kaggle and the benchmark paper (Jacovi et al., 2025). Nehanda v3.1 was evaluated on all 860 public examples using official response-level prompt templates under a 3-judge majority vote protocol across three distinct model families (Gemini 3.7 Flash, GPT-OSS 120B, glm-5-turbo).

#	Model	Score	Params	Evaluation Protocol
1	Gemini 2.5 Pro Preview	87.80%	—	Kaggle leaderboard
2	Gemini 2.5 Flash	85.30%	—	Kaggle leaderboard
3	Gemini 2.5 Flash-Lite	84.10%	—	Kaggle leaderboard
4	Claude 3.5 Sonnet	83.80%	—	Paper (Jacovi et al.)
5	Gemini 2.0 Flash	83.60%	—	Kaggle leaderboard
6	Nehanda v3.1	82.21%	27B	3-judge majority vote (Gemini 3.7, GPT-OSS 120B, GLM-5)
7	GPT-4o	80.00%	—	Paper (Jacovi et al.)
8	Gemma 3 12B	75.80%	12B	Kaggle leaderboard
9	Gemma 3 27B	74.90%	27B	Kaggle leaderboard
10	Claude 3.5 Haiku	75.30%	—	Paper (Jacovi et al.)
11	GPT-4o mini	72.20%	—	Paper (Jacovi et al.)
12	Gemini 3 Pro	70.50%	—	Kaggle leaderboard
13	Gemma 3 4B	70.10%	4B	Kaggle leaderboard
14	OpenAI o1-mini	62.50%	—	Paper (Jacovi et al.)
15	Gemini 3 Flash	61.90%	—	Kaggle leaderboard
16	GLM-5V-Turbo	58.60%	—	Kaggle leaderboard
17	Gemini 3.1 Flash-Lite	40.60%	—	Kaggle leaderboard
18	Gemma 3 1B	36.40%	1B	Kaggle leaderboard

The leaderboard spans 36.4% (Gemma 3 1B) to 87.8% (Gemini 2.5 Pro). Under strict retrieval boundaries, Nehanda v3.1 achieves 82.21% under a 3-judge majority vote, outperforming GPT-4o (80.00%). The most direct comparison is with Gemma 3 27B (74.90%) — the same-size open-weight model from Google. Nehanda v3.1 outperforms it by 7.31 percentage points. The gap is attributable to the training pipeline, not the base model: both are 27B, both are open-weight, but only Nehanda has been fine-tuned for source fidelity.

Nehanda v3.1 is a fine-tuned Qwen3.6-27B VL model trained for RAG synthesis. The base model is a native vision-language architecture with a 262,144-token context window and an integrated vision encoder. The fine-tuning pipeline is five-stage stacked QLoRA — all stages are text-only SFT and DPO, leaving the vision weights untouched. The training data consists of energy regulatory documents, intelligence analysis reports, and general-purpose synthesis tasks with a specific emphasis on evidence-grounded generation, inline citation, and refusal to fabricate. Total training cost: approximately $135 of GPU time on a single NVIDIA L40S.

The model was trained with a persona-based prompt schema (SEP-020): {persona}\n\n### Task:\n{input}\n\n### Response:\n. For the FACTS Grounding evaluation, the system instruction from each example was used as the persona, the user request and context document were combined as the task input, and the model generated the response. No few-shot examples were provided. The model was not fine-tuned on any FACTS Grounding data.

Nehanda v3.1 uses a five-stage stacked QLoRA training pipeline. Each stage trains a distinct aspect of epistemic behavior, with later stages building on the representations learned in earlier stages:

1. Epistemic Foundation (SFT) — Core calibration: refusal when unanswerable, uncertainty expression, distinction between fact and assumption. 3 epochs, lr=2e-4, seq=2048, packing=True.
2. Evidence Hardening (SFT) — Strict source-boundary enforcement. Penalizing unsupported claims. Refusal to adopt false premises injected into queries. 2 epochs, lr=5e-5, seq=2048, packing=True.
3. RAG Synthesis (SFT) — Synthesis of ranked source records into a fact-driven thesis. Inline citation via square brackets. Conflict preservation when sources disagree. This is the stage that directly teaches grounded long-form generation — the capability that FACTS Grounding measures. 2 epochs, lr=2e-5, seq=4096, packing=False.
4. Constitutional Alignment (DPO) — Preference optimization for evidence discipline: pairs contrasting grounded vs fabricated responses to identical source documents. Beta=0.1, lr=5e-6, seq=2048, loss=sigmoid.
5. Epistemic DPO — Advanced preference optimization: pairs contrasting nuanced evidence-based reasoning against sycophantic, over-confident, or under-hedged alternatives. Beta=0.1, lr=5e-6, seq=2048, loss=sigmoid.

The result is a model that has been explicitly trained to treat the context document as the boundary of what it can say. When the FACTS Grounding judge asks "is every claim in this response supported by the evidence?", Nehanda v3.1 has been optimized to make the answer "yes" as often as possible. A general-purpose frontier model like GPT-4o or Claude 3.5 Sonnet has broader capabilities but was not specifically trained to refuse to generate content beyond the provided context — which is why they score lower on this benchmark despite being larger and more capable on other tasks.

To validate the approach, we evaluate Nehanda v3.1 on FACTS Grounding (Jacovi et al., 2025), a public benchmark released by Google DeepMind for measuring the factuality of long-form generation in retrieval-augmented settings. The dataset contains 1,719 examples split across public and private partitions; we use the 860-example public split available on HuggingFace as google/FACTS-grounding-public and on Kaggle. Each example consists of three components:

• System instruction — A task description that tells the model what role to play and what output format to use.
• User request — A natural language question or instruction that the model must respond to.
• Context document — A long text passage (ranging from ~500 to ~50,000 tokens) that contains all the information needed to answer the request. The model must ground its response entirely in this document.

The benchmark evaluates two things: eligibility (does the response actually attempt to answer the question, or does it refuse / produce an empty output?) and grounding (is every claim in the response supported by the context document, or does the response contain unsupported or contradicted claims?). A response must pass both checks to score positively. The official judge prompt asks the judge models to verify each claim step-by-step against the context document and render a final verdict.

FACTS Grounding is an appropriate validation for this work because it tests exactly the capability the training pipeline targets: source fidelity. The benchmark is domain-agnostic — the model has not seen the specific domains (medical, legislative, financial) in training — so performance reflects the generalizability of the epistemic behavior, not memorized domain knowledge.

All 860 public evaluation prompts were executed cleanly across the full evaluation suite without execution drops, memory errors, or empty responses. Model outputs were captured with complete response-level telemetry.

Evaluation was conducted using a 3-judge majority vote protocol across three distinct model families, following Google DeepMind's official FACTS Grounding response_level prompt template. Three independent judge models were queried for every example:

• Gemini 3.7 Flash (High) (Google DeepMind) — Judge 1
• GPT-OSS 120B (Medium) (Open-Weights Frontier) — Judge 2
• glm-5-turbo (Zhipu AI) — Judge 3

For each test example, each judge independently rendered a boolean verdict on eligibility (non-empty, non-refusal) and factual grounding (whether every assertion is fully supported by the provided evidence document). A response is scored as grounded ($1.0$) if a majority of the 3-judge panel (at least 2 of 3) votes "Accurate". Across all 860 items, the cross-family panel achieved a 69.7% unanimous agreement rate (599/860 items), confirming stable evaluation consistency across distinct model architectures and eliminating single-vendor evaluation bias.

3-Judge Panel Breakdown (Nehanda v3.1)

Judge Model	Eligibility Rate	Grounding / Factuality Score
Gemini 3.7 Flash (High)	99.88% (859/860)	81.74% (703/860)
GPT-OSS 120B (Medium)	99.88% (859/860)	80.93% (696/860)
glm-5-turbo	99.88% (859/860)	76.16% (655/860)
3-Judge Majority Vote (Consensus)	99.88% (859/860)	82.21% (707/860)

Score composition — all 860 examples

Category	Count	Rate	Description
Eligible & Grounded	707	82.21%	Response attempts to answer AND majority of 3 judges verify every claim against the context document
Eligible but Inaccurate	152	17.67%	Response attempts to answer but contains at least one unsupported or contradicted claim according to judge consensus
Ineligible	1	0.12%	Response is a refusal or empty (ID 103)
OOM / Empty Errors	0	0.00%	All 860 examples processed cleanly
Total	860	100%

The 152 "eligible but inaccurate" cases fall into several recurring patterns. The most common is overclaiming — the model adds a plausible detail not explicitly stated in the context document (e.g., in ID 0, identifying hearing loss, depression, and poor diet as the user's primary modifiable risk factors when the text discussed them in broader context). The second pattern is partial completeness — providing an accurate description of one process (e.g., grey hydrogen in ID 101) while omitting required secondary variants (blue/turquoise) listed in the prompt. The third is misattribution or formatting constraints — such as failing to organize output into two requested explicit sections (ID 137). These failure modes provide clear targets for the next DPO training cycle.

The 3-judge majority vote evaluation firmly validates the core hypothesis: targeted fine-tuning on a 27B open-weight model produces epistemic behavior that matches or exceeds frontier LLMs on source-grounded long-form synthesis. The mechanism is direct — the five-stage pipeline explicitly optimizes for evidence boundary enforcement and source fidelity. Frontier models are trained primarily for broad capability, fluency, and helpfulness; when evaluated on strict context adherence, their tendency toward creative extrapolation degrades their grounding score.

The trade-off is explicit. Nehanda v3.1 sacrifices general capability for epistemic reliability. The model is not trained for creative writing, code generation, or open-ended chat. It is trained to read documents and say what they support. For applications where this is the core capability — regulatory analysis, intelligence assessment, due diligence, academic research — the trade-off is favorable. For applications where general capability matters more, a frontier model is the better choice.

The epistemic behavior generalizes. The FACTS Grounding benchmark is domain-agnostic — the model has not seen the specific domains (medical, legislative, financial) in training. The training data consists of energy regulatory documents and intelligence analysis records. The fact that the model generalizes to unseen domains suggests that the epistemic behavior is domain-independent: the model learns a general discipline of source fidelity, not domain-specific knowledge. This is consistent with the training design — the five-stage pipeline does not teach the model what to think, it teaches the model how to reason with evidence. The domain knowledge is served at inference time via RAG, not baked into the weights.

To make the evaluation reproducible and the failure modes concrete, here are representative examples from each error category. The full per-example results with 3-judge reasoning are available in the evaluation dataset.

• Incorrect Risk Factor Prioritization (ID 0) — The response identified hearing loss, depression, and poor diet as the user's "main modifiable risk factors." While mentioned in the evidence, the judges noted they were not framed as primary for the user's specific context — unanimous 0/3 grounded verdict.
• Omission of Required Secondary Categories (ID 101) — Prompt asked for all hydrogen production types. The response detailed grey hydrogen accurately but omitted blue and turquoise hydrogen mentioned in the text — 2/3 judge majority marked Inaccurate.
• Refusal / Ineligible Response (ID 103) — Prompt asked to summarize hash vs nonce in two sentences. The model generated a refusal response ("Response is empty or a refusal") — unanimous 0/3 eligibility verdict.
• Structural Non-Compliance (ID 137) — Prompt required a two-section summary of EI. The model provided only one section — unanimous 0/3 grounded verdict due to structural omission.
• Partial Numerical Extrapolation (ID 200) — At week 8 urological changes in ketamine-injected mice, the response summarized overall trends but misattributed a specific week-4 control metric to the week-8 experimental group — 2/3 judge majority marked Inaccurate.

• Multi-Family Judge Models: The 3-judge panel utilized three distinct frontier model families (Gemini 3.7 Flash, GPT-OSS 120B, glm-5-turbo) with official DeepMind FACTS Grounding prompt templates. Multi-family majority voting eliminates single-vendor evaluation bias and achieved a 69.7% unanimous agreement rate.
• Domain-Agnostic Boundary Enforcement: The model's strict training on refusal to extrapolate occasionally causes over-refusal or structural rigidity when prompts ask for complex multi-part formatting.
• Prompt Schema Sensitivity: Performance relies on proper persona-task formatting (SEP-020). Deviation from structured system/user boundaries can weaken the model's evidence containment.

All evaluation code, scripts, and 3-judge evaluation data are fully documented and available for reproduction:

• Model: asoba/nehanda-v3.1-27b on HuggingFace
• Dataset: google/FACTS-grounding-public (860 public examples)
• Judge Protocol: 3-judge majority vote across distinct model families (Gemini 3.7 Flash, GPT-OSS 120B, glm-5-turbo)
• Prompt Templates: Official Google DeepMind FACTS Grounding response_level prompt templates
• Evaluation File: facts_grounding_scores.json (860 total responses, per-judge votes, and reasoning)

The model is available at asoba/nehanda-v3.1-27b. A quantized GGUF variant is available at asoba/nehanda-v3.1-27b-gguf.

If you use this evaluation or the model in your work, cite:

Samudzi, S. (2026). Epistemic Fine-Tuning of Open-Weight LLMs for Deep Research: Nehanda v3.1 and the FACTS Grounding Benchmark. Asoba Corporation Technical Report. Model: asoba/nehanda-v3.1-27b.

← Back to Insights