rk500/Born-9B-Qwen3.5-9B-Preview

Born-9B Qwen3.5-9B Preview

Born-9B Preview is a PEFT LoRA adapter for Qwen/Qwen3.5-9B. It is trained for coding-agent behavior: concise planning, concrete code or patches, explicit checks, and a final user-facing result.

This release is the promoted Born-9B v2 adapter. Later recovery, hotfix, and preview-recovery experiments are documented in the project repo, but they are not promoted because they did not beat this adapter on the same local held-out gate.

Release Artifact

• —Hugging Face model ID: rk500/Born-9B-Qwen3.5-9B-Preview
• —Adapter type: PEFT LoRA adapter, not a merged full model.
• —Base model: Qwen/Qwen3.5-9B
• —Local project artifact: checkpoints/born-9b-v2-generated-expanded-lora
• —Training route: continuation from checkpoints/born-9b-v1-teichai-quick-lora, not a fresh adapter from base.
• —Intended use: coding-agent SFT behavior, repository repair planning, data-science debugging, tool-use closure, and explicit verification steps.
• —Not intended as: an official SWE-bench leaderboard claim, a general chatbot replacement, or a raw hidden chain-of-thought model.

Response Contract

Born-9B Preview was trained toward this visible response shape:

Plan:
- short concrete plan

Patch or Code:
<code, patch, commands, or exact action payload>

Checks:
- exact tests/checks to run

Result:
brief final user-facing result

The training target excludes raw hidden reasoning tags. Rows containing <think>, <thinking>, <reasoning>, or similar hidden-trace markers were filtered or sanitized before inclusion. The desired behavior is compact visible rationale: state invariants and decision rules, then produce the implementation and checks.

What Was Trained

The promoted v2 package is born9b_v2_generated_expanded.

Final mix report:

• —Total validated SFT rows: 7,097
• —Train / validation split: 6,672 / 425
• —Estimated tokens: 9.11M
• —Validation rejects: 0
• —Hidden tag rows in final validation: 0
• —Targeted curriculum rows: 85
• —Eval-derived DPO pairs available for analysis: 14
• —Duplicate prompts removed during v2 merge: 1,438
• —Max sequence length used for training: 4096

The v2 run intentionally mixed benchmark-style coding tasks, repair tasks, agent/tool traces, high-quality reasoning final answers, and a limited Irish-language supplement. The goal was not to maximize generic text volume; it was to push the model toward closure on coding-agent tasks.

Dataset Sources Used

The table below lists the sources recorded in the final v2 mix report. Row counts are the rows that survived into the v2 generated-expanded SFT mix after import, normalization, and deduplication.

External Dataset Names

The imported Hugging Face sources documented for the v1/v2 build include:

• —angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k
• —nohurry/Opus-4.6-Reasoning-3000x-filtered
• —sequelbox/Tachibana4-DeepSeek-V4-Pro
• —lambda/hermes-agent-reasoning-traces
• —Modotte/CodeX-2M-Thinking
• —Jackrong/DeepSeek-V4-Distill-8000x
• —DJLougen/hermes-agent-traces-filtered
• —open-thoughts/AgentTrove
• —Qwen/WebWorldData
• —Inferact/codex_swebenchpro_traces
• —TeichAI/claude-4.5-opus-high-reasoning-250x

These datasets were not imported as raw hidden chain-of-thought. The importer kept final answers, observable actions, code, patches, checks, and results. Rows were normalized into the Born response contract where possible.

Teacher And Synthetic Generation Sources

Teacher-generated and synthetic rows were produced across several provider lanes. The recorded teacher pool includes:

• —CrofAI: kimi-k2.6, mimo-v2.5-pro, deepseek-v4-pro, and attempted greg lanes.
• —OpenRouter: inclusionai/ring-2.6-1t, deepseek/deepseek-v4-flash, deepseek/deepseek-v4-pro, openrouter/owl-alpha, qwen/qwen3.6-plus, minimax variants, Arcee Trinity, Gemma, and Qwen thinking variants where available during the run.
• —Local/self generation: deterministic BigCodeBench-style rows, exact-code closure rows, data-science repair rows, agentic closure rows, and tool-use closure rows.

Provider rows were filtered for the expected section markers, minimum response quality, duplicate prompts, and hidden-tag leakage. Some generated lanes were rejected or stopped when schema compliance was poor.

Task Mix

Final v2 row counts by task kind:

Minor legacy task kinds with one or two rows are retained in the project reports but are not a meaningful part of the final behavior.

Training Configuration

Born-9B Preview v2 was trained with QLoRA SFT:

• —Model: Qwen/Qwen3.5-9B
• —Starting adapter: checkpoints/born-9b-v1-teichai-quick-lora
• —Output adapter: checkpoints/born-9b-v2-generated-expanded-lora
• —Quantization: 4-bit NF4
• —Torch dtype: bfloat16
• —Max sequence length: 4096
• —LoRA rank: 32
• —LoRA alpha: 64
• —LoRA dropout: 0.05
• —Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• —Epochs: 0.45
• —Optimizer steps: 188
• —Learning rate: 1.2e-5
• —Scheduler: cosine
• —Warmup ratio: 0.04
• —Effective batch size: 16
• —Weight decay: 0.01
• —Max grad norm: 0.3
• —Eval every 75 optimizer steps
• —Training GPU: NVIDIA A40 on RunPod

Final training telemetry:

• —Final train loss: 0.5978
• —Final eval loss: 0.7479
• —Final mean token accuracy: 0.8087
• —Final epoch: 0.4508
• —Completed steps: 188 / 188

Evaluation Snapshot

Fixed Local Born Self Eval

This is a project-local 25-task gate covering exact Python, repair, data science, agentic planning, and tool-use. It is not a public leaderboard.

The v2 release is promoted because it beats the base model and all later local recovery attempts on this same weighted gate.

SWE-bench Verified Proxy Sample

This is a 25-task issue-resolution proxy sample derived from SWE-bench Verified. It is not the official SWE-bench Docker harness.

HumanEval And MBPP Executable Slice

This is a fresh post-release A40 run over the first 25 local HumanEval rows and first 25 local MBPP rows. It executes Python assertions. It is still a small slice, not the full HumanEval/MBPP benchmark.

This is an honest regression on the combined exact-code slice: Born improves the MBPP sample but loses more on HumanEval. Treat exact-code improvement as future work, not a preview claim.

Later Attempts Not Promoted

Known Limitations

• —This is a LoRA adapter, so users need the Qwen/Qwen3.5-9B base model at inference time.
• —Evaluation is local and project-specific unless otherwise stated.
• —The SWE-bench result is a proxy sample, not the official Docker-based SWE-bench score.
• —The fresh HumanEval/MBPP executable slice trails base Qwen overall, so Born-9B Preview should not be marketed as a general exact-code benchmark win.
• —Tool-use score remains below base Qwen on the local suite, even though the total weighted score improves.
• —One known exact-code weakness in v2 is chunking strings as string slices instead of lists of characters.
• —Some public benchmark-style rows contributed to training-family coverage, so do not interpret training-adjacent probes as clean leaderboard evidence.
• —The model is optimized for coding-agent closure, not broad open-domain chat.

Loading

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

base_id = "Qwen/Qwen3.5-9B"
adapter_id = "rk500/Born-9B-Qwen3.5-9B-Preview"

tok = AutoTokenizer.from_pretrained(base_id, use_fast=True)
if tok.pad_token is None:
    tok.pad_token = tok.eos_token

quant = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

base = AutoModelForCausalLM.from_pretrained(
    base_id,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    quantization_config=quant,
)
model = PeftModel.from_pretrained(base, adapter_id)
model.eval()

Example generation:

messages = [
    {
        "role": "system",
        "content": (
            "You are Born-9B, a coding agent. Answer with Plan, Patch or Code, "
            "Checks, and Result. Be concise and complete."
        ),
    },
    {"role": "user", "content": "Fix this Python function and include tests: ..."},
]

text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(
        **inputs,
        max_new_tokens=900,
        do_sample=False,
        pad_token_id=tok.eos_token_id,
    )

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

Suggested Inference Prompt

You are Born-9B, a coding agent.
Answer with:
Plan:
- short concrete plan

Patch or Code:
the code, patch, or exact actions

Checks:
- exact tests/checks

Result:
final user-facing result

Do not expose hidden chain-of-thought. Be concise and complete.

Provenance Files In Project Repo

Key local files used to build and verify this release:

• —configs/distill-v2-generated-expanded.yaml
• —configs/lora-v2-eval-expanded.yaml
• —reports/born9b_v2_generated_expanded_mix_report.json
• —reports/born9b_v2_generated_expanded_validation.json
• —reports/born9b_v2_generated_local_born_self_25_report_corrected_2026_05_16.json
• —reports/qwen35_base_local_born_self_25_report_corrected_2026_05_16.json
• —reports/born9b_v2_swebench_verified_proxy_25_report.json
• —reports/qwen35_base_swebench_verified_proxy_25_report.json
• —docs/born-9b-v2-generation-log-2026-05-15.md
• —docs/born-9b-v2-runpod-training-status-2026-05-15.md
• —docs/hf-reasoning-agent-datasets-2026-05-14.md
• —docs/born-9b-preview-crof-recovery-2026-05-17.md

License

This adapter is released under Apache-2.0. The Qwen/Qwen3.5-9B model page lists Apache-2.0 at release time; users must comply with the base model license and with the licenses of any datasets they separately use for further training.

Citation

If you reference this preview artifact, cite it as:

@misc{born9b_qwen35_preview_2026,
  title        = {Born-9B Qwen3.5-9B Preview},
  author       = {LeemerLabs / Repath Khan},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/rk500/Born-9B-Qwen3.5-9B-Preview}},
  note         = {PEFT LoRA adapter for Qwen/Qwen3.5-9B}
}