prathmeshadsod/bondshift-organic-chemistry

BondShift: Organic Chemistry Mechanism Dataset

10,000 ground-truth-separated records for mechanism diagnosis, misconception repair, and chemistry tutoring.

A narrow, auditable V1 dataset built from independently constructed scenario blueprints and deterministic answer keys.

TL;DR

BondShift addresses the right answer, wrong mechanism problem. It trains models to examine electron flow, formal charge, intermediates, pathway choice, and stereochemical constraints instead of rewarding only a final product or reaction label.

The public release contains exactly 10,000 rows and preserves two views of each example: the audited source prompt/response, and Adaption's remastered enhanced_prompt/enhanced_completion. Approximately 80% of the source bank is mechanism core and 20% is broader science-bridge explanation.

Contents

• Problem statement
• Scope
• Dataset composition
• How each row was created
• Independent truth architecture
• Prompt design
• Response design
• What Adaption changed
• Quality control
• Adaption measurements
• Published schema
• Examples
• Loading the data
• Recommended use
• Limitations
• Provenance and reproducibility
• Citation

Problem statement

Many chemistry datasets supervise only a short answer, reaction name, or final product. That leaves several important failure modes weakly represented:

• an arrow begins from the wrong electron source;
• a mechanism creates an impossible or unjustified intermediate;
• a pathway is selected from one keyword while substrate and solvent evidence are ignored;
• a strong base is assumed to override anti-periplanar geometry;
• resonance contributors are described as separate molecules;
• a model gives an exact ratio or stereochemical label without enough information.

These errors matter in tutoring because a polished explanation can sound convincing while teaching the wrong causal model. BondShift was created to supervise the chain from visible evidence to a defensible mechanism and to teach the model when the evidence supports only a conditional answer.

The dataset is intentionally narrow. Six well-bounded V1 families were preferred over broad reaction coverage with unreliable ground truth.

Scope

BondShift V1 intentionally prefers reliable depth over broad reaction coverage.

Included chemistry families

Family	Rows	Share
Acid-base	1,654	16.54%
Resonance and formal charge	1,672	16.72%
SN1	1,676	16.76%
SN2	1,641	16.41%
E1	1,660	16.60%
E2	1,697	16.97%
Total	10,000	100%

Dedicated substitution/elimination competition and basic-stereochemistry template families were disabled for V1. Supported stereochemical reasoning appears only inside certified or restricted S_N2 and E2 scenarios.

Task taxonomy

Task type	Rows	Share
Predict outcome	2,088	20.88%
Mechanism debug	2,079	20.79%
Science bridge explanation	1,329	13.29%
Choose pathway	1,243	12.43%
Stereochemistry check	844	8.44%
Misconception check	836	8.36%
Explain key step	791	7.91%
Compare conditions	790	7.90%

Dataset composition

Tracks

Track	Purpose	Rows
bondshift_core	Direct mechanism analysis, debugging, outcomes, and pathway decisions	8,017
science_bridge	Explanations connecting symbolic mechanisms to charge flow, geometry, and observations	1,983

Difficulty

Difficulty	Rows	Share
Basic	2,015	20.15%
Intermediate	3,025	30.25%
Advanced	2,959	29.59%
Expert	2,001	20.01%

Difficulty describes the cognitive and explanatory demands of a row. It is separate from chemistry certainty. A harder question is not allowed to rely on a less reliable answer key merely to appear advanced.

Certification status

V1 status	Rows	Meaning
Certified	6,643	Supported by the V1 deterministic template rules
Restricted	3,357	Allowed only under bounded qualitative claims and additional validation rules
Disabled	0	Disabled templates are excluded from the final release

Restricted E1 and E2 templates avoid unsupported exact product ratios and overconfident condition-dependent claims.

How each row was created

The final dataset was selected from a deterministic 12,000-slot capability plan.

1. A seeded planner selects a chemistry family, task type, difficulty, and dataset track.
2. A curated V1 template constructs a prompt-safe scenario blueprint.
3. A deterministic compiler independently constructs the answer key and its provenance.
4. A prompt-author model turns only the visible blueprint into a natural user request.
5. Prompt validators reject answer leakage, artificial JSON, meta language, and unsupported requests.
6. A response-teacher model receives the frozen prompt and a factual grounding packet.
7. Response validators check answer-key alignment, internal leakage, restricted claims, and response quality.
8. Exact normalized duplicates are replaced before the immutable 10,000-row snapshot is written.

The accepted production release used nvidia/nemotron-3-ultra-550b-a55b as both the prompt-author and response-teacher model. These were separate stages with different inputs. No fallback teacher model was silently mixed into the accepted rows.

Independent truth architecture

No language model was allowed to invent the scenario facts, target answer, and validation truth as one self-confirming bundle.

Object	Contains	Visible to
Scenario blueprint	Substrate, reagents, medium, conditions, observations, and optional student claim	Prompt author
Answer key	Canonical mechanism, expected outcome, key decision, steps, bond changes, misconceptions, validation targets	Deterministic validators
Grounding packet	Factual mechanism data only; no ideal response prose	Response teacher
Natural prompt	The final user-facing question	Response teacher and SFT export
Natural response	Direct answer and relevant explanation	SFT export
MechanismIR-lite	Archival mechanism representation	Full audit export only

Every answer key records template ID, template version, rule-set version, generator version, validation level, manual-review requirement, and source type. This separation reduces self-confirming synthetic errors and makes a row traceable back to the deterministic rule that produced its target.

Prompt design

Prompts are intentionally varied rather than generated from one visible scaffold.

• short questions, contextual homework requests, misconception checks, and mechanism-debugging conversations;
• variable amounts of substrate, solvent, condition, and structural detail;
• direct and tentative user voices across all four difficulty levels;
• no mandatory JSON suffix, answer schema, or exposed validator target;
• prompts with insufficient information are expected to elicit a bounded or conditional response.

Across the frozen source snapshot, prompts average 80.62 whitespace-delimited words; the median is 83, the 25th and 75th percentiles are 55 and 103, and the observed range is 9 to 240 words.

Response design

Responses are natural assistant messages, not serialized answer keys. The target pattern is flexible:

1. answer the user's actual question;
2. identify the decisive mechanistic fact;
3. repair the misconception or invalid step when present;
4. add assumptions, caveats, or next steps only when they materially help.

The frozen source responses average 169.35 whitespace-delimited words; the median is 148, the 25th and 75th percentiles are 112 and 207, and the observed range is 32 to 653 words.

The source pipeline generated no separate private chain-of-thought. Its response field contains the complete user-facing explanation. The public Adaption export later added a distinct reasoning_trace column; that field is described separately below and must not be confused with the deterministic answer key or MechanismIR-lite.

Quality control

Final snapshot checks

The immutable release report records all of the following as passing:

Check	Result
Exactly 10,000 rows	Pass
Required full-metadata fields present	Pass
No empty prompts or responses	Pass
All snapshot validations accepted	Pass
No internal teacher leakage detected	Pass
Normalized prompts unique	Pass
Normalized responses unique	Pass
Stable slot IDs and ordinals unique	Pass
Every row matches the frozen run ID	Pass

Repair and deduplication

• 56 records identified by the pre-export audit were regenerated; all 56 passed the final repair audit.
• 59 residual output wrappers were stripped in the final transform.
• Eight exact normalized prompt duplicates were replaced; no exact prompt or response duplicate remains.
• Three resonance/formal-charge rows received targeted scientific corrections after review.
• The final report records zero unresolved upload blockers.

These checks establish schema validity, template consistency, leakage controls, and exact-deduplication status. They do not establish exhaustive chemical correctness for every possible use.

The checks above apply to the frozen source snapshot. The public repository contains Adaption-remastered columns in addition to the source pair, so users combining or transforming the enhanced text should perform a fresh duplicate and chemistry audit for their exact training export.

What Adaption changed

Adaption preserved the original chemistry pair and added a remastered view. The enhancements commonly make the task more explicit, request clearer deliverables, organize the completion into a more pedagogical explanation, and separate facts, conditions, misconceptions, and conclusions more visibly.

This is more than a numerical quality-score change. It affects tone, structure, instruction-following, and how the chemical argument is communicated. The original pair remains valuable for studying natural questions and the source pipeline; the enhanced pair represents the adapted training experience.

Adaption measurements

Measure	Original	Adaptive	Change
Quality score	8.0	9.5	+18.8% relative
Grade	B	A	B to A
Percentile	17.8	57.7	+39.9 points

Adaption reports the uploaded dataset as 10,000 English rows, with Science and Academic-education as its detected domains. The table reproduces platform-displayed measurements. No public item-level rubric, confidence interval, or independent chemistry benchmark was available, so these values should not be interpreted as a universal accuracy score.

Published schema

The Hugging Face dataset currently exposes five string columns:

Field	Type	Description
prompt	string	Original natural user request from the audited source snapshot
response	string	Original validated, user-facing source answer
enhanced_prompt	string	Adaption-remastered user instruction
enhanced_completion	string	Adaption-remastered assistant target
reasoning_trace	string	Auxiliary reasoning data created during adaptation

Which columns should I train on?

Goal	User field	Assistant field
Reproduce the Adaption-enhanced instruction behavior	enhanced_prompt	enhanced_completion
Train on the compact audited source pair	prompt	response
Compare adaptation effects	Compare source and enhanced pairs side by side	Do not merge them into one answer

reasoning_trace is not the source answer key, not MechanismIR-lite, and not required for ordinary SFT. Do not expose it as a user-visible answer or silently concatenate it into the assistant target. Any experiment that trains on it should be treated as a separate, explicitly evaluated variant.

Full audit export

The internal full-metadata JSONL additionally retains:

Field group	Examples
Stable identity	slot_id, ordinal, run_id, content_hash
Capability labels	dataset_track, difficulty, family, task_type
Generation lineage	prompt-author model, response-teacher model, raw-attempt paths
Validation	prompt and response validation records, V1 certification status
Scientific provenance	answer-key provenance and MechanismIR-lite
Snapshot lineage	final transform and snapshot metadata

MechanismIR-lite and answer-key provenance are sidecar audit data. They are not assistant responses and should not be used as private chain-of-thought targets.

Examples

Mechanism debugging

Prompt: I'm trying to draw the mechanism for bromomethane with NaOH in DMSO. I showed the bromide leaving first to give a methyl carbocation, then hydroxide attacking. Something feels off because there's nothing to stabilize that carbocation, but I'm not sure what the actual problem is or how to fix it.

Response: Your instinct is exactly right: a methyl carbocation is far too unstable to form under these conditions. The reaction proceeds by a concerted S_N2 pathway: hydroxide attacks the carbon from the backside while the C-Br bond breaks in a single step. There is no carbocation intermediate. Because the carbon in bromomethane is not a stereocenter, you will not observe inversion of configuration. The product is simply methanol and bromide ion. DMSO is a polar aprotic solvent, which keeps hydroxide highly nucleophilic and favors this bimolecular displacement. Redraw the mechanism with one curved arrow from hydroxide to carbon and a second from the C-Br bond to bromide, both in one step.

Resonance explanation

Prompt: Why are both C-O bonds the same length in acetate if one is a double bond and one is a single bond in the resonance structures?

Response: The two drawings are contributors, not distinct molecules that acetate flips between. The real structure is a resonance hybrid in which the pi electrons and negative charge are delocalized over both oxygens. Each C-O bond therefore has an average bond order of 1.5 and the two bond lengths are identical, intermediate between typical single and double C-O bonds. The total -1 charge is conserved and shared across the two oxygens.

Loading the data

Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset(
    "prathmeshadsod/bondshift-organic-chemistry",
    split="train",
)

print(dataset)
print(dataset.column_names)
print(dataset[0])

Build the adapted chat pair

def to_messages(row):
    return {
        "messages": [
            {"role": "user", "content": row["enhanced_prompt"]},
            {"role": "assistant", "content": row["enhanced_completion"]},
        ]
    }

chat_dataset = dataset.map(to_messages)
print(chat_dataset[0]["messages"])

Use prompt and response instead when the experiment specifically targets the frozen source behavior.

Recommended use

• supervised fine-tuning for organic chemistry tutoring;
• evaluation-set construction after independent review and strict train/eval separation;
• studying mechanism misconceptions and explanation styles;
• controlled experiments on narrow-domain adaptation.

When constructing an evaluation set, split by deterministic template lineage or scenario signature rather than random row alone. This reduces near-variant leakage between training and evaluation.

Limitations

• Coverage is deliberately limited to six V1 families. The dataset does not certify carbonyl addition/substitution, electrophilic addition, rearrangements, complex EAS, radical, pericyclic, organometallic, or broad redox mechanisms.
• Inputs and outputs are synthetic. Natural language was produced by one large response model and may carry stylistic or factual biases from that model.
• Automated checks and deterministic answer keys reduce risk but are not a substitute for row-level expert review.
• Exact normalized duplicates are absent; semantic similarity can still exist because multiple rows teach the same core chemical rule from different user contexts.
• Text-only prompts do not validate image interpretation, hand-drawn structures, atom mapping, or arbitrary SMILES.
• Condition-sensitive selectivity and product ratios are represented qualitatively. Exact quantitative claims should not be inferred from this dataset.
• Some advanced prompts are difficult in presentation, not broader in certified chemistry. Difficulty must not be read as a certainty score.
• The dataset is educational and must not be used as laboratory, industrial, clinical, or safety guidance.

Provenance and reproducibility

Artifact	Value
Snapshot version	bondshift.final_snapshot.v1
Run ID	bondshift_production_10k_nim_seed_20260730_v3
Planner seed	20260730
Created	2026-07-30 15:21:11 UTC
Rows	10,000
Prompt author	nvidia/nemotron-3-ultra-550b-a55b
Response teacher	nvidia/nemotron-3-ultra-550b-a55b

Release file	SHA-256
Public adapted data/train_0.jsonl	9f5b906b9544ce71790c7326654481a82ddb11ef4412b703a95aec37a231dd79
bondshift_adaption_prompt_response_10000_20260730.jsonl	6b7ff73ccb2c0ee29e4f8e5315eb9794b80d0302b033ea26d962aadde7fe7129
bondshift_adaption_prompt_response_10000_20260730.csv	bf5d73e99f24684c04c986557df574fb979be1902562295f09293860efe9a271
bondshift_adaption_chat_10000_20260730.jsonl	5aed12abf05b5b6db6ad6da210e6124a005a850a9be73ef8b1186f549898722c
bondshift_complete_with_metadata_10000_20260730.jsonl	b5d21bc8a6483ef54280fb971679c3c46a1130fd9174a605563dcea7b2aa3259

The public 201 MB train_0.jsonl is the Adaption-remastered five-column dataset. The smaller prompt-response JSONL is the frozen source pair, the chat JSONL is its chat-formatted equivalent, and the full-metadata JSONL is the scientific audit artifact.

Citation

@misc{adsod2026bondshiftdataset,
  author = {Prathmesh Adsod},
  title = {BondShift Organic Chemistry Dataset},
  year = {2026},
  note = {10,000-row mechanism-reasoning dataset for the Adaption AutoScientist Challenge}
}

Acknowledgements

Created for the Adaption AutoScientist Challenge. Adaption supplied the adaptation and measurement workflow used for the displayed quality results. The dataset construction pipeline kept deterministic chemistry truth separate from natural prompt and response generation.