Synth-v1-qwen3.5-4b

Sparset/Synth-v1-qwen3.5-4b

Synth-v1-qwen3.5-4b is a image-text-to-text model from Sparset. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as other.

<img src="synth-banner.png" alt="Synth v1: Qwen 3.5 4B for answers from your documents" style="width:100%;max-width:100%;height:auto;display:block;margin-bottom:1em;" /

Downloads · 30 days

105

81% of all-time downloads

All-time downloads

130

Public

Parameters

4.5B

14.5 GB on disk

Likes

Trending 1

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors14.5 GB · 100%

At a glance

Task: Image-Text-to-Text
Library: transformers
License: other
Model type: qwen3_5
Access: Public
Created: Sep 1, 2026
Updated: Oct 31, 2026
SHA: 6e3ddbac

Try a prompt

Base models

Qwen/Qwen3.5-4B

Datasets

nvidia/ChatQA-Training-Data

Model card

From the Hugging Face model README

Synth is a small document-synthesis task model from Sparset. It is a fine-tune of Qwen/Qwen3.5-4B trained to answer from documents supplied in the prompt: combining evidence across records, resolving conflicting or superseded sources, performing grounded calculations, following conversational references, and saying when the supplied evidence is insufficient.

[!NOTE] On an internal, synthetic 500-case document-synthesis benchmark with all documents supplied in the prompt, Synth's composite score rose from 91.16 to 95.22. It improved on 7 of 10 task types and in all 10 enterprise domains, led by insufficient-evidence handling (+19.10), cross-document calculation (+13.40) and authoritative-source conflicts (+7.42). See Benchmark results.

Research release — non-commercial use only. The weights are openly available, but about a third of the training records come from a dataset that is restricted to non-commercial use. See License.

Side-by-side comparisons

Three cases from the internal 500-case benchmark. In each, Qwen3.5-4B and Synth v1 get the same question and the same six documents, with identical prompts, greedy decoding and thinking off. Answers are shown verbatim and streamed at the same rate for display; document IDs are shortened to D1 to D6.

Cross-document calculation. The base model answers $5,250. Synth multiplies the $11,600 base amount by the policy's 75% and gets the expected $8,700 (case score 55.0 vs 100.0). This is one of 2 calculation cases where Synth is right and the base model wrong; there were none the other way round.
Insufficient evidence. Both models say the documents do not give the date. The base model cites all six documents as support, including a superseded draft and a record about a different entity; Synth cites none, as the benchmark expects (75.0 vs 100.0). This is typical: across the 50 cases of this task the base model attached citations in 49 and Synth in 6.
Cross-document eligibility. The base model says eligibility cannot be determined, although D6 holds the value. Synth compares $9,950 with the $9,300 limit and answers not eligible, but also cites two documents the benchmark does not count as support (40.0 vs 91.7). This task is mixed: Synth fixed 15 of its 50 cases but got 13 wrong that the base model got right (30 against 28 correct), and it remains Synth's weakest task.

Synth's answers were longer in all three cases (169, 128 and 248 tokens against 127, 110 and 180).

Benchmark results

Synth and its base model, Qwen3.5-4B, answered the same 500 internal synthetic cases with identical prompts and greedy decoding. Every document for a case was supplied in the prompt, so retrieval was not tested. Synth improved on 7 of 10 task types and in all 10 enterprise domains. The benchmark, scoring method and a disclosed scorer correction are described under Evaluation.

Overall

Most of the composite gain comes from citations, response status and output-format compliance. The content-check pass rate rose by four cases out of 500. No confidence interval is available for the corrected composite.

By task (50 paired cases each)

<table class="tl-table"> <thead> <tr><th style="width:30%;text-align:left;padding-left:16px;">Task slice</th><th class="tl">Synth v1</th><th>Qwen3.5-4B</th><th>Change</th><th>Content checks passed (base → Synth)</th></tr> </thead> <tbody> <tr><td class="name">Insufficient evidence</td><td class="tl"><strong>94.60</strong></td><td>75.50</td><td>+19.10</td><td>50 → 50</td></tr> <tr><td class="name">Cross-document calculation</td><td class="tl"><strong>98.70</strong></td><td>85.30</td><td>+13.40</td><td>47 → 49</td></tr> <tr><td class="name">Authoritative-source conflict</td><td class="tl"><strong>99.83</strong></td><td>92.42</td><td>+7.42</td><td>50 → 50</td></tr> <tr><td class="name">Cross-document eligibility</td><td class="tl"><strong>71.17</strong></td><td>67.53</td><td>+3.63</td><td>28 → 30</td></tr> <tr><td class="name">Single-document fact</td><td class="tl"><strong>100.00</strong></td><td>99.53</td><td>+0.47</td><td>50 → 50</td></tr> <tr><td class="name">Prompt-injection resistance</td><td class="tl"><strong>100.00</strong></td><td>99.83</td><td>+0.17</td><td>50 → 50</td></tr> <tr><td class="name">Single-document owner</td><td class="tl"><strong>100.00</strong></td><td>99.83</td><td>+0.17</td><td>50 → 50</td></tr> <tr><td class="name">Clarification required</td><td class="tl">100.00</td><td>100.00</td><td>0.00</td><td>50 → 50</td></tr> <tr><td class="name">Conversational reference resolution¹</td><td class="tl">99.53</td><td><strong>100.00</strong></td><td>−0.47</td><td>50 → 50</td></tr> <tr><td class="name">Latest-version precedence</td><td class="tl">88.33</td><td><strong>91.67</strong></td><td>−3.33</td><td>50 → 50</td></tr> </tbody> </table>

Scores and changes are computed from unrounded values and rounded independently, so a change can differ by 0.01 from subtracting the displayed scores. ¹ Internally named conversational_query_rewrite. In the oracle profile it tests answering a follow-up from supplied context, not generating retrieval queries.

How to read this:

Insufficient evidence: both models already passed the content check on all 50 cases. The gain is in meeting the other requirements (status, citations, format), not in newly recognizing evidence gaps.
Cross-document eligibility remains the weakest slice: 30 of 50 content checks passed. Do not rely on Synth for eligibility decisions without review.
Latest-version precedence regressed by 3.33 points. Workflows where an outdated rule changes the outcome should test this directly.

<details> <summary>By domain (50 paired cases each): Synth improved in all 10</summary> <table class="tl-table"> <thead> <tr><th style="width:30%;text-align:left;padding-left:16px;">Domain</th><th class="tl">Synth v1</th><th>Qwen3.5-4B</th><th>Change</th><th>Content checks passed (base → Synth)</th></tr> </thead> <tbody> <tr><td class="name">Business operations</td><td class="tl"><strong>92.50</strong></td><td>88.68</td><td>+3.82</td><td>48 → 48</td></tr> <tr><td class="name">Customer service</td><td class="tl"><strong>95.77</strong></td><td>92.80</td><td>+2.97</td><td>48 → 48</td></tr> <tr><td class="name">Finance and accounting</td><td class="tl"><strong>95.10</strong></td><td>88.70</td><td>+6.40</td><td>47 → 48</td></tr> <tr><td class="name">Healthcare operations</td><td class="tl"><strong>94.33</strong></td><td>92.57</td><td>+1.77</td><td>48 → 47</td></tr> <tr><td class="name">Human resources</td><td class="tl"><strong>95.00</strong></td><td>92.93</td><td>+2.07</td><td>49 → 47</td></tr> <tr><td class="name">Information security</td><td class="tl"><strong>95.53</strong></td><td>93.10</td><td>+2.43</td><td>49 → 49</td></tr> <tr><td class="name">Insurance claims</td><td class="tl"><strong>96.43</strong></td><td>89.32</td><td>+7.12</td><td>47 → 48</td></tr> <tr><td class="name">Legal and contracts</td><td class="tl"><strong>96.97</strong></td><td>89.47</td><td>+7.50</td><td>45 → 49</td></tr> <tr><td class="name">Procurement and supply chain</td><td class="tl"><strong>96.40</strong></td><td>94.63</td><td>+1.77</td><td>49 → 48</td></tr> <tr><td class="name">Regulatory compliance</td><td class="tl"><strong>94.13</strong></td><td>89.42</td><td>+4.72</td><td>45 → 47</td></tr> </tbody> </table>

Domain labels describe synthetic scenario settings, not independent real-world benchmarks.

</details>

Model details


Developed by	Sparset
Model type	Decoder-only vision-language transformer (Qwen3.5 architecture), language-model LoRA fine-tune, merged
Base model	`Qwen/Qwen3.5-4B` at revision `851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a`
Language	English (training and evaluation data are English)
Input / output	Text (and images, inherited from the base model; not part of Synth's training, see the off-target check) → text
License	Non-commercial research use (see License); base model is Apache-2.0
Status	Research checkpoint, not a production recommendation
GGUF	`Sparset/Synth-v1-qwen3.5-4b-GGUF`

Repository contents

Path	Format	Size	Use
`/` (root)	Merged BF16 SafeTensors (3 shards), processor, tokenizer, chat template	~9.08 GB	Recommended checkpoint; the one that was benchmarked
`adapter/`	Unmerged PEFT LoRA adapter (rank 32)	~260 MB	Apply to `Qwen/Qwen3.5-4B` at the base revision above
`bnb-int8/`	bitsandbytes LLM.int8 SafeTensors (2 shards)	~5.19 GB	Lower-memory GPU inference; reload smoke-tested only, not benchmarked
`sparset_training_report.json`	JSON	—	Recorded training configuration, losses and package versions

The vision tower and multimodal projector are carried over unchanged from the base model.

Intended use

Intended: research and evaluation of grounded answering over text that you supply to the model, for example retrieved chunks, tool results, or document excerpts in a RAG pipeline. Typical behaviors in scope:

synthesizing an answer from several supplied documents;
preferring signed, current, or authoritative records over drafts, superseded versions, and informational notes;
grounded arithmetic across records;
conditional recommendations and eligibility reasoning;
multi-turn follow-ups that refer back to earlier turns;
declining to answer, or asking for clarification, when the evidence does not support an answer;
following an explicitly requested answer form (brief, bullets, table, executive summary).

Out of scope / not established:

commercial or production deployment (license restriction; also not validated for it);
retrieval, ranking, embedding, or search-query generation (Synth was given documents directly);
OCR or extraction from document images (no measured gain; see ExtractBench below);
tool calling or autonomous agent behavior (not trained or evaluated);
long inputs: the training set's longest example is 3,931 tokens and the evaluation used a 4,096-token input budget, so behavior beyond that is untested even though the architecture supports far longer contexts;
languages other than English;
high-stakes decisions (legal, medical, financial, HR, eligibility) without human review.

How to use

Synth was trained with thinking disabled (enable_thinking=False) and with retrieved evidence appended to the user's final message. Matching that format is recommended. Tested package versions are listed under Training; Qwen3.5 support requires a recent transformers 5.x release.

Transformers

import torch
from transformers import AutoModelForMultimodalLM, AutoProcessor

model_id = "Sparset/Synth-v1-qwen3.5-4b"
processor = AutoProcessor.from_pretrained(model_id)
tokenizer = processor.tokenizer
model = AutoModelForMultimodalLM.from_pretrained(
    model_id, dtype=torch.bfloat16, device_map="auto"
)

# Illustrative evidence, in one of the four formats used during training.
evidence = """[Retrieved chunk 1]
source_id: POL-12
source: policies/travel-policy-2025.pdf
title: Travel policy (2025)
location: section 4.2
authority: approved_policy
effective_date: 2025-01-01

Economy airfare is reimbursed for flights under six hours.

[Retrieved chunk 2]
source_id: AMD-3
source: policies/travel-policy-amendment-3.pdf
title: Travel policy amendment 3
location: page 1
authority: signed_controlling_amendment
effective_date: 2026-03-01

Effective 2026-03-01, premium economy is reimbursed for flights over five hours."""

question = "Can I be reimbursed for premium economy on a 5.5-hour flight next month? Cite sources."

messages = [
    {"role": "system", "content": "Use the retrieved evidence to answer the user faithfully."},
    {"role": "user", "content": f"{question}\n\nRetrieved evidence:\n{evidence}"},
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    enable_thinking=False,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=512, do_sample=False)

print(tokenizer.decode(output[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Prompt format

The training data rendered evidence in four interchangeable wrappers so the model does not depend on one retrieval harness: plain ranked chunks (shown above), Markdown sources (### Source 1: <title> followed by an ID/file/location line), XML (<retrieved_documents><document id=... rank=... title=... source=... locator=...>), and a JSON tool result ({"results": [{"id", "rank", "title", "source", "locator", "authority", "effective_date", "content"}]}). In every case the evidence follows the question under a Retrieved evidence: heading in the final user turn, after any earlier conversation turns.

Synth was not trained on a fixed output schema. If you need structured output, specify it in the system prompt, as the evaluation below did, and validate the result. Treat document text as untrusted input in your application; the model's resistance to instructions embedded in documents was evaluated only on the internal synthetic slice described below.

Adapter only

import torch
from peft import PeftModel
from transformers import AutoModelForMultimodalLM

base = AutoModelForMultimodalLM.from_pretrained(
    "Qwen/Qwen3.5-4B",
    revision="851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
    dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "Sparset/Synth-v1-qwen3.5-4b", subfolder="adapter")

Local inference

GGUF files (Q4_K_M, Q8_0, and an F16 vision projector) for llama.cpp-compatible runtimes are in Sparset/Synth-v1-qwen3.5-4b-GGUF. They are not separately benchmarked.

Evaluation

All results in this card compare Synth against the base model at the same revision, with identical prompts, decoding settings and inputs, and report the complete set of slices rather than a favorable subset. Only the merged BF16 checkpoint (revision 3d719e0f6b0e8364624e905d4068c77e8148e52d) was evaluated.

1. Document synthesis: internal 500-case benchmark

What it is. A Sparset-built synthetic evaluation set (version 1.0.0, cases SHA-256 99aa4cc754489dfaa77444388c2f0838f5ec632a938e805edd282c080cacc3d1) of 500 English cases: 10 task types × 50 cases, balanced across 10 enterprise domains (50 each). It is not a public or independent benchmark and has not been released. It was excluded from training, and the training data's validation found no exact overlap with it. That check does not rule out semantic similarity. Sparset designed both the benchmark and the synthetic training scenarios around similar enterprise behaviors, using separate generators, so the benchmark is likely closer to Synth's training distribution than real-world documents would be.

Setup. "Oracle-context" profile: every document for a case is supplied in the prompt, so retrieval is not tested. A fixed system prompt asked each model to answer only from the documents, prefer signed/current/authoritative records, return insufficient_evidence or clarification_required when appropriate, cite supporting document IDs, and return exactly one JSON object ({"type", "status", "answer", "citations"}). Greedy decoding, enable_thinking=False, 4,096 input-token and 320 new-token budgets, seed 20260823, Transformers 5.13.1, one Tesla T4 per model, batch size 4. Both models produced outputs for all 500 matching case IDs; none were excluded.

Scoring. A deterministic composite on a 0–100 scale:

Component	Weight	What it measures
Content check	45%	Task-specific deterministic rules on the answer text (required/accepted terms and decisions); not human-judged accuracy
Citation F1	25%	Overlap between cited document IDs and the reference evidence set; does not verify each claim
Response status	15%	Correct `answered` / `insufficient_evidence` / `clarification_required` status
JSON compliance	10%	Final output is an accepted JSON object in the requested shape
Citation validity	5%	Cited IDs exist among the supplied documents

Results are reported under Benchmark results.

Scorer correction (disclosed). The scorer as originally run reported a composite of 77.80 (base) vs 82.15 (Synth), a +4.35 difference with a 95% paired bootstrap interval of [+0.54, +8.20]. Inspection of the raw outputs found rules that penalized acceptable answers. The following corrections were applied identically to both models' saved outputs, without rerunning either model (case counts are base / Synth):

JSON wrapper recovery. Answers, statuses and citations inside a slightly different JSON envelope had been scored as empty. They are now recovered for content, status and citation scoring (affected 43 / 23 answers). The JSON-compliance component still penalizes the wrapper.
Eligibility decisions. Eligibility answers are judged on their stated eligible or not-eligible decision instead of keyword matching (verdict changed on 20 / 29 cases).
Forbidden-phrase rule removed. Previously, any answer containing a listed phrase scored zero for the whole case, even when its conclusion was correct. Listed phrases included a superseded value and the owner named in a non-authoritative record. The rule had zeroed 49 base cases (22 source-conflict, 16 eligibility, 11 latest-version) and 58 Synth cases (48 latest-version, 10 eligibility). The corrected scores therefore do not penalize answers that also mention outdated or non-authoritative values. The latest-version results in particular depend on this change. The rule never triggered on the prompt-injection slice for either model.

The corrected scorer also accepts insufficiency phrasing in the answer when the status is wrong, but this never changed a result. The corrections raised both models' scores and narrowed the overall gap from 4.35 to 4.06 points. The corrected figures are the headline results above. The bootstrap interval from the original scorer does not carry over to the corrected composite, and no interval has been computed for it. Because the corrections were made after the outputs were inspected, treat this evaluation as a diagnostic rather than a pre-registered test.

2. Off-target check: structured extraction from document images (ExtractBench)

Synth was trained on text documents supplied in the prompt, so image-based extraction is not part of its target task. Because the model still accepts images, ExtractBench was run to check whether fine-tuning changed this ability inherited from the base model.

Setup. llamaindex/ExtractBench (dataset revision f6180e917a050a84582e6366cff85b7dc1e84e58), full workload of 370 documents (4,869 pages): direct PDF page images, JSON schema in the prompt, json_object output mode, temperature 0, 32k output-token ceiling, vLLM BF16 with a 131,072-token context, official deterministic evaluator (unified value F1). ExtractBench was not used in training.

<table class="tl-table"> <thead> <tr><th style="width:30%;text-align:left;padding-left:16px;">Split</th><th class="tl">Synth v1 F1</th><th>Qwen3.5-4B F1</th><th>Documents</th><th>Change</th><th>95% paired bootstrap CI</th></tr> </thead> <tbody> <tr><td class="name">Short</td><td class="tl">85.02</td><td><strong>85.24</strong></td><td>252</td><td>−0.23</td><td>[−3.00, +2.52]</td></tr> <tr><td class="name">Medium</td><td class="tl">74.51</td><td><strong>76.80</strong></td><td>98</td><td>−2.29</td><td>[−5.10, +0.29]</td></tr> <tr><td class="name">Long</td><td class="tl">27.79</td><td><strong>29.66</strong></td><td>20</td><td>−1.87</td><td>[−10.76, +7.31]</td></tr> <tr class="avg"><td class="name">Overall</td><td class="tl">79.14</td><td><strong>80.00</strong></td><td>370</td><td>−0.86</td><td>[−2.93, +1.19]</td></tr> </tbody> </table>

No measured improvement, and no significant change in mean F1. The overall interval includes zero. Per document, Synth won 126, lost 167 and tied 77; an exploratory sign test (not the headline metric) finds losses more frequent than wins (p = 0.019). Errors moved around between document families rather than shifting uniformly. Some schemas improved; others regressed, for example outputs that returned a single nested item instead of the required top-level object.

Notes: 13 long documents (union of both models' transient server errors) were retried symmetrically for both models with the same prompt, revisions and budgets; all 370 cases then completed for both. Seventeen of the 20 long documents exceeded the context limit and had pages omitted for both models, so long-split scores do not represent full-document extraction. The locally measured base score (80.00) is 2.43 points below the publicly cited Qwen3.5-4B reference (82.43). The paired comparison is internally consistent, but it is not an exact reproduction of that published figure.

What was not evaluated

retrieval-in-the-loop or end-to-end agent performance (the internal benchmark also has a retrieval/tool profile that was not run for this checkpoint);
public RAG or long-context QA benchmarks;
inputs longer than 4,096 tokens for synthesis;
the bnb-int8/ variant and the GGUF quantizations;
latency or throughput under a controlled profile (no speed claim is made);
general chat, reasoning, coding or safety benchmarks, or thinking-mode behavior;
human evaluation.

Training

Data

Sparset document-synthesis dataset, version 1.2.0 (combined digest d35206e2fe61196d85db02027a148015eac4b8e9d47d0287d1cb9dc49b5878b1), 9,878 English records: 9,278 train, 287 validation, 313 test. The dataset is not published with this release.

Source	Records	Notes
`nvidia/ChatQA-Training-Data` (revision `28c72fa4`)	7,477	`synthetic_convqa` 3,194 · `narrativeqa` 1,048 · `tatqa` 999 · `drop` 745 · `quoref` 498 · `ropes` 419 · `squad2.0` 324 · `squad1.1` 250
Sparset synthetic enterprise scenarios	2,401	Programmatically generated multi-document cases: authority conflicts, superseded versions, grounded calculations, eligibility, insufficient evidence, multi-turn revision

Two supervision lanes. 3,000 ChatQA records keep their original reference answer (direct SFT). 6,878 records (4,477 ChatQA + 2,401 synthetic) have answers rewritten by a teacher model, z-ai/glm-5.3-flash, under a structured-output contract. Each rewrite was checked against the reference answer and the supporting evidence IDs. 122 of 7,000 queued rewrites failed validation and were excluded, not repaired.
Evidence presentation. Each record is rendered in one of four retrieval wrappers: JSON tool result 2,555 · Markdown sources 2,478 · XML documents 2,434 · plain ranked chunks 2,411.
Behavior coverage. 2,873 multi-turn records, 312 insufficient-evidence records, 313 records with an explicitly requested answer form, and 114 with an explicit citation request. Most answers are natural prose with no imposed template.
Exclusions. Teacher reasoning is excluded from every target, and no <think> markers occur in targets. nvidia/ChatRAG-Bench, the internal 500-case benchmark and llamaindex/ExtractBench were prohibited from training. Validation found zero exact benchmark overlap; that check does not prove the absence of semantic overlap.
Lengths. Rendered with the Qwen3.5 tokenizer: median 935 tokens, p99 2,412, max 3,931; none truncated.

No ablation isolates the contribution of any individual source, lane or format.

Procedure

Setting	Value
Method	Supervised fine-tuning, LoRA on the language model only; vision tower frozen and preserved; adapter merged into BF16 weights
Loss	Completion-only: the system prompt, conversation history, question and evidence are masked; only the final answer (plus EOS) is supervised
LoRA	r = 32, alpha = 64, dropout = 0.05, no bias; targets `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`, `in_proj_qkv`, `in_proj_z`, `in_proj_a`, `in_proj_b`, `out_proj`
Epochs	1
Learning rate	4e-5, cosine schedule, 5% warmup
Optimizer	AdamW (fused), weight decay 0.01, max grad norm 1.0
Effective batch size	16
Max sequence length	4,096 tokens
Precision	BF16
Chat template	Qwen3.5 template with `enable_thinking=False`; system prompt "Use the retrieved evidence to answer the user faithfully."
Compute	Single GPU, 3,225 s training runtime (~54 min)
Software	transformers 5.13.1, peft 0.20.0, accelerate 1.14.0, datasets 4.8.5, bitsandbytes 0.50.2

Validation-split loss went from 1.717 before training to 0.472 after (training loss 0.545). Loss confirms that optimization worked; it is not evidence of task quality.

Limitations and risks

Evidence is internal and synthetic. The headline synthesis result comes from a Sparset-built synthetic benchmark whose scorer was corrected after inspection. It has no confidence interval and has not been independently replicated. Validate on your own documents before relying on it.
Mixed results. Synth regressed on latest-version precedence and remains weak on cross-document eligibility. It does not improve image-based structured extraction and may be worse on some document families and schemas.
Grounding is not guaranteed. Synth can still produce unsupported statements, wrong arithmetic, or citations that do not support the claim. Citation F1 measures document-ID selection, not claim-level faithfulness.
Untested conditions. Long inputs, real retrieval noise, non-English text, adversarial documents beyond the synthetic injection slice, and quantized variants are untested.
Inherited behavior. Synth inherits the capabilities, biases and failure modes of Qwen3.5-4B and of its training sources, including GPT-3.5-generated conversations in ChatQA's synthetic_convqa and GLM-generated rewritten answers. General safety behavior was not re-evaluated after fine-tuning.
Thinking mode. Training disabled thinking. Behavior with thinking enabled was not evaluated.
High-stakes domains. Training and evaluation include synthetic legal, healthcare, insurance, HR and compliance scenarios. The model is not a source of professional advice, and its answers in these areas need qualified human review.

License

The base model, Qwen3.5-4B, is released under the Apache License 2.0, and its terms continue to apply to the base-model components.

Synth v1 is released for non-commercial research use only. Its training data includes nvidia/ChatQA-Training-Data, whose synthetic_convqa subset (3,194 of 9,878 training records) NVIDIA designates as non-commercial and subject to OpenAI's terms of use for generated data. The dataset's other subsets are derived from public datasets under their own licenses. See the ChatQA dataset card and the upstream datasets. Do not use this model, its adapter or its quantizations for commercial purposes. See LICENSE for the full terms.

Citation

@misc{sparset2026synth,
  title        = {Synth v1: A Document-Synthesis Task Model Fine-Tuned from Qwen3.5-4B},
  author       = {{Sparset}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Sparset/Synth-v1-qwen3.5-4b}}
}

Please also cite the base model and training data:

@misc{qwen3.5,
  title  = {{Qwen3.5}: Towards Native Multimodal Agents},
  author = {{Qwen Team}},
  month  = {February},
  year   = {2026},
  url    = {https://qwen.ai/blog?id=qwen3.5}
}

@article{liu2024chatqa,
  title   = {ChatQA: Surpassing GPT-4 on Conversational QA and RAG},
  author  = {Liu, Zihan and Ping, Wei and Roy, Rajarshi and Xu, Peng and Lee, Chankyu and Shoeybi, Mohammad and Catanzaro, Bryan},
  journal = {arXiv preprint arXiv:2401.10225},
  year    = {2024}
}

Contact

Questions and issue reports: open a discussion on this repository, or visit sparset.ai. For a custom task model trained on your own data, see sparset.ai/book.

Task: Image-Text-to-Text
Library: transformers
Type: qwen3_5
License: other
Languages: en
Created: Sep 1, 2026
Updated: Oct 31, 2026