README.md

---
base_model: togethercomputer/Meta-Llama-3.3-70B-Instruct-Reference
base_model_relation: adapter
language:
  - en
library_name: peft
pipeline_tag: text-generation
datasets:
  - prathmeshadsod/bondshift-organic-chemistry
tags:
  - chemistry
  - organic-chemistry
  - mechanism-reasoning
  - education
  - sft
  - lora
  - adaption-autoscientist
---

<div align="center" style="padding:46px 30px;border-radius:30px;background:linear-gradient(135deg,#ffffff 0%,#f8fbff 18%,#eef2ff 42%,#f4e8ff 66%,#e6fbff 100%);border:1px solid #c7d2fe;box-shadow:0 18px 46px rgba(30,64,175,0.13);">

<div style="margin-bottom:14px;">
  <span style="display:inline-block;width:15px;height:15px;border-radius:999px;background:#dbeafe;margin:0 5px;border:1px solid #93c5fd;"></span>
  <span style="display:inline-block;width:15px;height:15px;border-radius:999px;background:#c4b5fd;margin:0 5px;border:1px solid #a78bfa;"></span>
  <span style="display:inline-block;width:15px;height:15px;border-radius:999px;background:#a5b4fc;margin:0 5px;border:1px solid #818cf8;"></span>
  <span style="display:inline-block;width:15px;height:15px;border-radius:999px;background:#a5f3fc;margin:0 5px;border:1px solid #67e8f9;"></span>
</div>

<h1 style="margin:0 0 12px 0;">BondShift: Organic Mechanism Reasoning</h1>

<p style="font-size:19px;margin:0 0 12px 0;"><strong>A mechanism tutor that diagnoses why an answer is right or wrong, not merely which product appears at the end.</strong></p>

<p style="max-width:920px;margin:0 auto;">LoRA-adapted from Llama 3.3 70B Instruct with a 10,000-row, ground-truth-separated training set.</p>

</div>

<p align="center"><img src="assets/bondshift-win-rate.png" alt="BondShift evaluation results" width="100%" style="border-radius:22px;border:1px solid #c7d2fe;box-shadow:0 14px 34px rgba(30,64,175,0.10);background:linear-gradient(180deg,#ffffff,#f8fbff);"></p>

## TL;DR

<div style="padding:22px 24px;border-radius:22px;background:linear-gradient(135deg,#ffffff 0%,#eef5ff 32%,#f5ecff 67%,#ecfeff 100%);border:1px solid #c7d2fe;box-shadow:0 10px 26px rgba(30,64,175,0.07);">

BondShift targets a common failure in organic-chemistry assistance: producing a plausible product while giving an
invalid electron-flow story. It connects reagents and conditions to electron movement, charge, intermediates,
stereochemical constraints, pathway choice, and the final outcome.

The target behavior is diagnostic. BondShift should identify the first invalid step, explain the controlling chemical
principle, repair the mechanism, and state what additional structural or condition information is needed when the answer
is genuinely underdetermined.

The submitted model is a **LoRA adapter**, not a standalone 70B checkpoint. It must be used with the exact base-model
family recorded in the AutoScientist configuration.

</div>

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## Contents

- [Why BondShift](#why-bondshift)
- [The design response](#the-design-response)
- [Supported capabilities](#supported-capabilities)
- [Example](#example)
- [Evaluation](#evaluation)
- [Model details](#model-details)
- [Training configuration](#training-configuration)
- [Training data](#training-data)
- [What Adaption improved](#what-adaption-improved)
- [Ground-truth architecture](#ground-truth-architecture)
- [How to use](#how-to-use)
- [Intended use](#intended-use)
- [Limitations](#limitations)
- [Reproducibility](#reproducibility)
- [Citation](#citation)

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## Why BondShift

<div style="padding:20px 22px;border-radius:22px;background:linear-gradient(135deg,#ffffff,#f8fbff,#eef2ff);border:1px solid #c7d2fe;box-shadow:0 8px 22px rgba(30,64,175,0.06);">

Organic chemistry is not solved by reaction-name recall alone. A student can memorize "strong base means E2" and still
draw an impossible arrow, remove the wrong beta hydrogen, miss an anti-periplanar requirement, create an unstable
carbocation, or treat resonance contributors as rapidly interconverting molecules.

These are high-value tutoring failures because the final product can look correct even when the reasoning that produced
it is not. Conventional answer-only data rewards the destination and may never teach the model to locate the broken
step.

BondShift therefore trains the reasoning layer between problem and conclusion:

| Failure mode | Desired BondShift behavior |
|---|---|
| Correct product, invalid mechanism | Find and repair the first chemically invalid step |
| Mechanism chosen from one keyword | Weigh substrate, nucleophile/base, solvent, geometry, and conditions together |
| Strong reagent treated as overriding structure | Explain the geometric or orbital constraint that still applies |
| Missing structure or conditions | Give a bounded answer or ask for the decisive missing fact |
| Resonance or charge misconception | Track electron and charge conservation explicitly |
| Unsupported selectivity claim | Avoid false precision and state the decision-changing conditions |

</div>

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## The design response

<div style="padding:20px 22px;border-radius:22px;background:linear-gradient(135deg,#ffffff,#f8f5ff,#f3e8ff);border:1px solid #ddd6fe;box-shadow:0 8px 22px rgba(88,28,135,0.06);">

BondShift was built around independently constructed chemistry answer keys instead of allowing one language model to
invent the question, answer, and validation truth together. The response target is still natural tutoring prose: answer
first, explain the decisive mechanism, correct the misconception, and add caveats only when they matter.

The production design also deliberately restricted V1 to chemistry families whose ground truth could be represented
reliably. Breadth was postponed when it would have required fragile or overconfident answer keys.

</div>

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## Supported capabilities

| Capability | Examples |
|---|---|
| Mechanism debugging | Find an invalid arrow, impossible intermediate, or wrong leaving-group step |
| Outcome reasoning | Predict a qualitative product or pathway from a supported substrate and conditions |
| Key-step explanation | Explain proton transfer, ionization, backside attack, or beta elimination |
| Pathway choice | Compare supported S<sub>N</sub>1, S<sub>N</sub>2, E1, and E2 pathways |
| Stereochemical checks | Explain inversion and cyclohexane trans-diaxial requirements when the structure is sufficiently specified |
| Misconception repair | Correct common errors about charge, resonance, nucleophilicity, basicity, and carbocations |
| Science bridge explanations | Connect arrow pushing to charge flow, orbital alignment, and observable chemical behavior |

The validated V1 training scope contains six families: **acid-base**, **resonance/formal charge**, **S<sub>N</sub>1**,
**S<sub>N</sub>2**, **E1**, and **E2**.

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## Example

This is an actual prompt-response pair from the frozen training snapshot.

> **User**
>
> I'm looking at this E2 with bromocyclohexane and NaOEt/EtOH heat. My notes say any beta-H can go because ethoxide is
> strong, but the problem hints the leaving group has to be axial and a chair flip might be needed. Does that mean only
> certain hydrogens actually work, or does the strong base override the geometry thing?

> **BondShift**
>
> In an E2 reaction, the base strength doesn't override the geometric requirement. The reaction is concerted and
> demands that the breaking C-H and C-Br bonds be anti-periplanar so their sigma orbitals can align with the forming pi
> bond. In a cyclohexane chair, this alignment is only achieved when the leaving group (Br) and a beta-hydrogen are both
> axial, i.e. trans-diaxial.
>
> If bromocyclohexane starts with Br equatorial, a chair flip is needed to place Br axial. Only then can an axial
> beta-hydrogen on an adjacent carbon participate. Hydrogens that are equatorial, or axial but not anti to the leaving
> group, cannot react because the orbital overlap is wrong. So, despite ethoxide being a strong base, only the
> trans-diaxial hydrogens are geometrically competent, and the product is cyclohexene.

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## Evaluation

<a id="autoscientist-win-rates"></a>
<details open style="margin:16px 0;border:1px solid #c7d2fe;border-radius:18px;padding:0 16px;background:linear-gradient(135deg,#ffffff,#f8fbff,#f7f2ff);">
<summary style="cursor:pointer;padding:14px 0;"><strong>AutoScientist win rates</strong></summary>

| Evaluation slice | Base model | Adapted model |
|---|---:|---:|
| AutoScientist evaluation on this dataset | 3 | **98** |
| Science category evaluation | 14 | **86** |

These are the whole-number win-rate labels displayed by the Adaption AutoScientist evaluation interface for training
experiment `c7f0a1b0-8286-4387-8f08-f2e0a4b74998`. The interface did not expose sample counts, confidence intervals,
or a public item-level evaluation set. The values should therefore be read as platform-reported preference results,
not as universal chemistry accuracy estimates. The own-dataset labels may also reflect display rounding.

</details>

<a id="dataset-adaptation-signals"></a>
<details open style="margin:16px 0;border:1px solid #ddd6fe;border-radius:18px;padding:0 16px;background:linear-gradient(135deg,#ffffff,#f8fbff,#f7f2ff);">
<summary style="cursor:pointer;padding:14px 0;"><strong>Dataset adaptation signals</strong></summary>

The dataset used for this run was also measured before and after Adaption processing:

| Measure | Before | After |
|---|---:|---:|
| Quality score | 8.0 | **9.5** |
| Grade | B | **A** |
| Percentile | 17.8 | **57.7** |

The platform reports the quality-score change as **18.8% relative improvement**. These are platform measurements of
the submitted dataset, not independent chemistry benchmarks.

The improvement is visible beyond the score. The adapted records add clearer task framing, more explicit deliverables,
better organized explanations, and stronger misconception-focused teaching while preserving the chemical problem being
solved.

<p align="center"><img src="assets/bondshift-training.png" alt="BondShift training telemetry" width="100%" style="border-radius:22px;border:1px solid #ddd6fe;box-shadow:0 14px 34px rgba(88,28,135,0.09);background:linear-gradient(180deg,#ffffff,#faf7ff);"></p>

Training curves document the run's optimization telemetry. They do not, by themselves, establish chemical correctness
or out-of-distribution generalization.

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

</details>

## Model details

| Field | Verified value |
|---|---|
| Release type | LoRA adapter |
| Training method | Supervised fine-tuning (SFT) |
| Data format | Chat |
| Base model recorded by AutoScientist | `meta-llama/Llama-3.3-70B-Instruct-Reference` |
| Base path recorded by exported adapter | `togethercomputer/Meta-Llama-3.3-70B-Instruct-Reference` |
| Base model size | 70B |
| AutoScientist training model label | `adaption_llama_3_3_70b_instru_bondshift_organic_chemis_c7f0a1b0` |
| Trained model name | `adaption_organic_mechanism_qa` |
| Training experiment ID | `c7f0a1b0-8286-4387-8f08-f2e0a4b74998` |
| Fine-tune job ID | `a49eb748-1696-466a-8755-9c2952376891` |

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## Training configuration

The AutoScientist-selected configuration was used unchanged.

| Parameter | Value |
|---|---:|
| Epochs | 3 |
| Batch size | `max` |
| Evaluations | 5 |
| Learning rate | `1e-4` |
| Scheduler | Cosine |
| Scheduler cycles | 0.5 |
| Warmup ratio | 0.03 |
| Minimum LR ratio | 0.1 |
| Weight decay | 0 |
| Max gradient norm | 2 |
| LoRA rank | 32 |
| LoRA alpha | 64 |
| LoRA dropout | 0 |
| Trainable modules | `all-linear` |
| Train on inputs | `false` |

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## Training data

BondShift was trained from exactly **10,000 English chemistry records**. The public dataset exposes the audited source
pair and the Adaption-remastered pair side by side:

| Slice | Rows |
|---|---:|
| BondShift mechanism core | 8,017 |
| Science bridge | 1,983 |
| Total | **10,000** |

The final mix is approximately 20% basic, 30% intermediate, 30% advanced, and 20% expert. All six V1 chemistry
families are represented near evenly. Prompt styles range from short questions to contextual debugging requests, while
assistant responses use flexible natural prose rather than one fixed answer template.

Every row earns its place: the final 10,000 were selected from a deterministic 12,000-slot candidate plan, checked for
scope and answer-key alignment, screened for internal leakage, and deduplicated before the immutable release snapshot
was created.

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## What Adaption improved

<div style="padding:20px 22px;border-radius:22px;background:linear-gradient(135deg,#f8fbff,#eef2ff,#f5ecff,#ecfeff);border:1px solid #c7d2fe;box-shadow:0 10px 26px rgba(67,56,202,0.07);">

The source corpus already supplied natural questions, grounded answers, and strict mechanism validation. Adaption then
added a second, enhanced view of every record:

| Dataset field | Role |
|---|---|
| `prompt` | Original natural user question |
| `response` | Original validated answer |
| `enhanced_prompt` | Adaption-remastered instruction with clearer task framing |
| `enhanced_completion` | Adaption-remastered teaching response |
| `reasoning_trace` | Auxiliary platform-generated reasoning data |

The adapted AutoScientist run is associated with the enhanced instruction/completion view. The auxiliary reasoning
trace is not BondShift's deterministic answer key and is not the archival MechanismIR sidecar.

</div>

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## Ground-truth architecture

<div style="padding:20px 22px;border-radius:22px;background:linear-gradient(135deg,#ffffff,#eef2ff,#ecfeff);border:1px solid #bae6fd;box-shadow:0 10px 26px rgba(8,145,178,0.06);">

The data pipeline deliberately separates five concerns:

1. A deterministic **scenario blueprint** contains only facts that may be shown to the prompt author.
2. A separately constructed **answer key** stores the expected mechanism, outcome, bond changes, misconception target,
   validation targets, and provenance.
3. The prompt author receives the blueprint, never the hidden answer key.
4. The response teacher receives the frozen user prompt plus a factual grounding packet, never ideal response prose.
5. Validators independently check structure, leakage, restricted-template claims, and alignment with the answer key.

The training response contains the user-facing explanation and conclusion. No private chain-of-thought was generated or
exported. MechanismIR-lite is archival validation metadata and is not forced into the model's natural-language answer.

</div>

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## How to use

<a id="adaption-interface"></a>
<details open style="margin:16px 0;border:1px solid #bae6fd;border-radius:18px;padding:0 16px;background:linear-gradient(135deg,#ffffff,#f8fbff,#f7f2ff);">
<summary style="cursor:pointer;padding:14px 0;"><strong>Adaption interface</strong></summary>

Open dataset ID `9e740edb-5d61-49c0-9510-b37919676e4a` in Adaption, select **Interfaces**, and use the BondShift
organic-chemistry mechanism companion. A strong query includes the substrate, reagents, solvent or medium, conditions,
and the exact step or conclusion that is confusing.

</details>

<a id="local-adapter-inference"></a>
<details open style="margin:16px 0;border:1px solid #c4b5fd;border-radius:18px;padding:0 16px;background:linear-gradient(135deg,#ffffff,#f8fbff,#f7f2ff);">
<summary style="cursor:pointer;padding:14px 0;"><strong>Local adapter inference</strong></summary>

This release is a 1.66 GB LoRA adapter and cannot be loaded as a standalone causal language model. The repository now
contains the adapter weights, adapter configuration, tokenizer, and chat template. PEFT can read the exact base path
from `adapter_config.json`.

```python
import torch
from peft import PeftConfig, PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

adapter_id = "prathmeshadsod/BondShift-Llama-3.3-70B-Instruct"
peft_config = PeftConfig.from_pretrained(adapter_id)
base_model_path = peft_config.base_model_name_or_path

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base = AutoModelForCausalLM.from_pretrained(
    base_model_path,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, adapter_id)

messages = [{
    "role": "user",
    "content": (
        "Why must bromocyclohexane have an axial leaving group before an E2 "
        "elimination can occur?"
    ),
}]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=384, do_sample=False)

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

The exported configuration currently records `togethercomputer/Meta-Llama-3.3-70B-Instruct-Reference`. Access to a
compatible base checkpoint and hardware capable of serving a 70B model are still required. Quantization and adapter
merging should be tested separately; this repository is not a merged full checkpoint.

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

</details>

## Intended use

BondShift is intended for:

- undergraduate mechanism tutoring and formative feedback;
- explaining supported arrow-pushing decisions;
- debugging a student's proposed mechanism;
- generating study examples within the validated V1 scope;
- research on natural-language chemistry tutoring.

It is not intended to replace an instructor, validate a synthesis, provide laboratory safety instructions, or support
clinical, industrial, or high-stakes chemical decisions.

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## Limitations

- V1 covers acid-base, resonance/formal charge, S<sub>N</sub>1, S<sub>N</sub>2, E1, and E2. Carbonyl chemistry,
  electrophilic addition, rearrangement-heavy chemistry, complex aromatic substitution, radical, pericyclic,
  organometallic, and broad oxidation/reduction mechanisms were not certified for this release.
- The model consumes text. It was not validated as an image, sketch, molecular-graph, or SMILES interpretation system.
- Stereochemical labels require complete structural and CIP information. Ambiguous prompts should receive a conditional
  answer or a request for the missing structure.
- Product ratios and pathway dominance can depend on concentration, temperature, solvent, and substrate detail. The V1
  data intentionally avoids unsupported exact ratios, but the model can still overstate a qualitative preference.
- Deterministic template consistency and automated answer-key checks reduce errors; they do not prove exhaustive chemical
  correctness. Expert review remains appropriate for consequential use.
- Synthetic prompt and teacher styles can transfer into the adapted model. Performance may fall on unfamiliar notation,
  advanced reaction families, or adversarially incomplete questions.
- The reported win rates come from the Adaption interface. No confidence intervals or public item-level evaluation set
  were available for independent statistical analysis.

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## Reproducibility

| Artifact | Identifier |
|---|---|
| Frozen dataset snapshot | `bondshift.final_snapshot.v1` |
| Snapshot run ID | `bondshift_production_10k_nim_seed_20260730_v3` |
| Snapshot date | 2026-07-30 |
| Planner seed | 20260730 |
| Prompt-author model | `nvidia/nemotron-3-ultra-550b-a55b` |
| Response-teacher model | `nvidia/nemotron-3-ultra-550b-a55b` |
| Prompt-response JSONL SHA-256 | `6b7ff73ccb2c0ee29e4f8e5315eb9794b80d0302b033ea26d962aadde7fe7129` |

Only one prompt-author model and one response-teacher model were used in the accepted production release. Raw provider
attempts, validation records, deterministic provenance, repair reports, and the full metadata snapshot were retained for
auditability.

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## Citation

```bibtex
@misc{adsod2026bondshift,
  author = {Prathmesh Adsod},
  title = {BondShift Organic Chemistry: Mechanism-Reasoning LoRA},
  year = {2026},
  note = {Adaption AutoScientist Challenge submission}
}
```

<div style="height:3px;border-radius:999px;background:linear-gradient(90deg,#e0e7ff 0%,#93c5fd 22%,#818cf8 48%,#c084fc 70%,#67e8f9 88%,#f8fafc 100%);margin:32px 0 20px 0;"></div>

## Acknowledgements

Built for the Adaption AutoScientist Challenge. Adaption provided dataset adaptation, training configuration selection,
LoRA training, and the displayed evaluation results. The production data pipeline used deterministic chemistry templates
and a large reasoning model for natural prompt and response realization, with the truth and generation paths kept
separate.