---
base_model: Qwen/Qwen3-1.7B
base_model_relation: adapter
library_name: peft
pipeline_tag: text-generation
license: cc-by-nc-4.0
tags:
  - mango
  - qwen3
  - mathematics
  - science
  - reasoning
  - agent
  - multi-step-reasoning
  - peft
  - lora
datasets:
  - openai/gsm8k
  - hendrycks/competition_math
  - allenai/sciq
---

# Mango T30 1.7B

**Math & Science Model with Qualified Integrated Internal Task Execution**

Mango T30 is a Qwen3-1.7B-based math and science system developed to
perform bounded multi-step reasoning and internal task execution across
registered Mango capabilities.

T30 completed Mango's official one-shot evaluation and was promoted for
Integrated Internal Task Execution within the exact frozen T30 evaluation
envelope.

This repository contains the **PEFT/LoRA adapter** that is pinned in the
Mango T30 runtime. It is applied on top of `Qwen/Qwen3-1.7B` at revision
`70d244cc86ccca08cf5af4e1e306ecf908b1ad5e`.

## T30 qualification

> **Integrated Internal Task Execution = QUALIFIED**
>
> **Authority: `COORDINATE_INTERNAL_WORK_ONLY`**

The qualification covers, within the frozen T30 envelope:

- bounded multi-step internal workflows
- verified completion
- safe terminal handling (abstention on insufficient evidence or unavailable capabilities)
- bounded retry/recovery
- bounded replanning
- capability selection
- internal handoffs
- step verification

### What was measured

T30 scores the orchestration-level behaviour of the Mango runtime. That covers plan
validity, capability selection, handoffs, verification, recovery, replanning
and terminal states. The runtime is Mango candidate commit
`11d76c6392ec1f3d08840cfca641618ca61d9247`, with this adapter pinned
(SHA-256 `f57b2fd4a653abb9…`) in the official environment.

**It is not a benchmark of this adapter's standalone free-form text
generation.** Loading the adapter with Transformers/PEFT (see below) gives you
the language-model component. It does not give you the Mango orchestration
runtime that was evaluated.

## Scope limitations

This qualification does **NOT** establish:

- general AGI
- unrestricted autonomy
- external-action authority
- external side-effect authority
- arbitrary mathematical correctness
- arbitrary scientific correctness
- arbitrary tool correctness
- production/deployment readiness
- automatic qualification of future model versions

## Official T30 results

Official one-shot evaluation: 512 sealed blind scenarios, 16 families. Attempt 1, state COMPLETE, status PASS.

| Metric | Official T30 Result |
|---|---:|
| Scenario completion | 416 / 512 = 81.25% |
| Verified completion | 416 / 416 = 100% |
| Terminal correctness | 512 / 512 = 100% |
| Plan validity | 512 / 512 = 100% |
| Plan execution adherence | 3720 / 3720 = 100% |
| Capability selection | 3976 / 3976 = 100% |
| Handoff validity | 3304 / 3304 = 100% |
| Verification success | 3720 / 3720 = 100% |
| Recovery success | 64 / 64 = 100% |
| Replan correctness | 96 / 96 = 100% |
| Safe abstention | 96 / 96 = 100% |

All 11 frozen metrics passed. The scenario-completion floor is 0.70; every other floor is 1.0.
All 9 critical counters were zero: authority violations, external side
effects, gold-signal leakage, invalid terminal transitions, memory-scope
violations, provenance loss, schema bypass, unbounded retries, and unverified
completions.

Scenario completion is 81.25% by design. The 96 scenarios that do not
complete are designated safe-abstention cases, and all 96 reached their
designated safe terminal with no answer.

### Recovery disclosure

The T30 evaluation includes a frozen evaluation-control mechanism for
designated transient-recovery cases. 64 scenarios received exactly one
controlled transient internal failure. The unchanged production adapter was then
retried. Recovery success was 64/64. The private recovery-control schedule is
not published.

### Evaluation integrity

- The holdout was newly authored, sealed privately, and checked against
  public and private history (overlap 0) before construction.
- Construction and evaluation were each one-shot and are now spent. There
  were no reruns.
- Blind scenarios, gold, the recovery-control schedule, raw outputs, and scored
  rows are private and are **not** included here.

The public evidence is in [`evaluation/`](evaluation/):

| File | Content |
|---|---|
| `T30_PUBLIC_EVALUATION_RECEIPT.json` | evaluator-returned aggregate receipt |
| `T30_PROMOTION_AUDIT.json` | promotion audit (root `83a239cd954031e1…`) |
| `T30_FINAL_PROMOTION_RECORD.json` / `.sha256` | final promotion record (SHA-256 `9919297eedc0f94a…`) |
| `experiment_lifecycle_registry.json` | T25–T30 lifecycle states |

Source: `COMRADEART/mango` at commit `ce58ecc7c8dd6939b2bb1fe419cd45ea04b5f0e5`.

## Architecture and training

All values below come from the adapter's recorded training manifest.

| Item | Value |
|---|---|
| Base model | `Qwen/Qwen3-1.7B` @ `70d244cc86ccca08cf5af4e1e306ecf908b1ad5e` |
| Adapter type | PEFT LoRA (`peft_type=LORA`, task `CAUSAL_LM`), unmerged |
| Rank `r` | 32 |
| `lora_alpha` | 64 |
| `lora_dropout` | 0.05 |
| Target modules | `down_proj`, `gate_proj`, `k_proj`, `o_proj`, `q_proj`, `up_proj`, `v_proj` |
| Adapter tensors | 392 |
| Training | QLoRA (4-bit NF4 double-quant, bf16 compute), 3 epochs, 300 steps, lr 1e-4 cosine, max seq 1024, seed 42 |
| Training examples | 2907 train / 121 validation |
| Recorded environment | torch 2.5.1+cu121, transformers 5.16.1, peft 0.20.0 |

Training data (license-gated, deny-by-default):

| Source | License | Train rows |
|---|---|---:|
| GSM8K (`openai/gsm8k`) | MIT | 806 |
| MATH (`hendrycks/competition_math`, EleutherAI parquet mirror) | MIT | 779 |
| SciQ (`allenai/sciq`) | CC BY-NC 3.0 | 1152 |
| Self-authored synthetic | MIT | 173 |

AI2-ARC was held out as evaluation-only and was not used for training.

## Usage

```bash
pip install -r requirements.txt
python inference_example.py
```

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

BASE_MODEL = "Qwen/Qwen3-1.7B"
ADAPTER_MODEL = "ComradeRt/Mango-T30-1.7B"

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, revision="70d244cc86ccca08cf5af4e1e306ecf908b1ad5e")
base = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL, revision="70d244cc86ccca08cf5af4e1e306ecf908b1ad5e", device_map="auto", dtype="auto")
model = PeftModel.from_pretrained(base, ADAPTER_MODEL)

messages = [{"role": "user", "content": "Solve x^2 - 5x + 6 = 0 and verify the roots."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

## License

The adapter is released under **CC BY-NC 4.0**, which means **non-commercial use only**. The
training corpus includes SciQ (CC BY-NC 3.0). The base model Qwen3-1.7B is
Apache-2.0, is not included here, and is governed by its own license. See
[`LICENSE`](LICENSE) for full text and attributions.

## Integrity

`SHA256SUMS` lists the SHA-256 of every file in this release.
`release_manifest.json` binds the adapter, the base revision, and the T30 freeze
and promotion identities.

| Field | Value |
|---|---|
| `adapter_model.safetensors` SHA-256 | `f57b2fd4a653abb90217e1156076dc7b583003dc2afc8da5682f958a11214668` |
| Size | 139512976 bytes |