LoRA adapter fine-tuned on Qwen3-4B via ORPO (Odds Ratio Preference Optimization) for B2B sales email generation. Trained on preference pairs targeting five failure modes identified in a Week 10 production audit of the Tenacious sales agent.

Evaluation Results

Evaluated on the sealed held-out split of Tenacious-Bench v0.1 (n=32 tasks).

Arm	Mean Score / 5.0	95% CI
Week 10 baseline (base model)	4.008	[3.668, 4.301]
Prompt-engineered baseline	4.172	[3.887, 4.449]
This adapter (ORPO)	4.462	[4.193, 4.704]

Delta	Δ	p-value	Result
vs Week 10 baseline (Delta A)	+0.454	0.001	✓ PASS
vs Prompt-engineered (Delta B)	+0.290	0.021	✓ PASS

Gain by failure mode:

Failure mode	Δ
tone_drift (38% of Week 10 failures)	+0.679
signal_missing (29%)	+0.393
formulaic (8%)	+0.250
trajectory (21%)	+0.154

Model Details

Property	Value
Base model	`unsloth/Qwen3-4B-bnb-4bit`
Training method	ORPO — Odds Ratio Preference Optimization (Hong et al., 2024)
LoRA rank r	32
LoRA alpha α	32
Trainable parameters	66,060,288 (1.62% of 4.09B total)
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training precision	fp16 (T4 GPU)

Training Hyperparameters

python

beta            = 0.1    # ORPO odds-ratio weight λ
learning_rate   = 5e-5
epochs          = 5
batch_size      = 2      # per device
grad_accum      = 8      # effective batch = 16
max_length      = 512
max_prompt_length = 256
optimizer       = "adamw_8bit"
lr_scheduler    = "cosine"
warmup_steps    = 10
seed            = 42

Training Data

381 preference pairs generated from 127 training tasks in Tenacious-Bench v0.1.

Chosen: Gemini Flash-generated outputs following all Tenacious style rules
Rejected: Gemini Flash-generated outputs deliberately violating the task's failure_mode_tag
Token overlap (chosen vs rejected): avg 16.8% (range 4.8–26.8%)
Format: TRL canonical — prompt = formatted prompt, chosen/rejected = response text only

Note: Rejected outputs are Gemini-synthesized failure-mode injections, not captured production failures. Cross-family leakage prevention was not applied (Gemini both generates and judges). See dataset card for details.

Intended Use

Generating B2B sales emails, follow-ups, and objection responses for Tenacious-style outbound workflows. Specifically optimized for:

Signal-grounded opening lines (references verifiable prospect trigger)
Zero banned phrases (47-phrase prohibited list)
Single calendar CTA ([CALENDLY_LINK])
Word-count compliance
No pricing on first touch

How to Use

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="Chalie-lijalem/tenacious-orpo-qwen3-4b",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

messages = [
    {"role": "system", "content": "You are a B2B sales assistant for Tenacious. Write direct, signal-led sales messages with no banned phrases."},
    {"role": "user",   "content": "Write an email_outreach for this prospect.\n\nContext:\nProspect: Sarah Chen, VP of Revenue at Lattice. Series C $45M 2023-Q2. LinkedIn post 3 days ago about sales rep ramp time.\n\nConstraints:\n- under 120 words\n- include [CALENDLY_LINK]\n\nWrite only the message body."},
]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

import torch
with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=256, temperature=0.0, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Limitations

English-only. Tested on US B2B SaaS contexts only.
Optimized for Tenacious-specific banned phrase list and style rules. May not generalize to other sales brands.
Evaluation dataset is small (n=32 held-out tasks). Confidence intervals are wide.
trajectory failure mode shows smallest improvement (Δ=+0.154) — multi-turn consistency remains partially unresolved.
Rejected preference pairs are synthetic (Gemini-generated) rather than captured from production failures.

References

Hong et al. (2024). ORPO: Monolithic Preference Optimization without Reference Model. https://arxiv.org/abs/2403.07691
Dataset: Chalie-lijalem/tenacious-bench-v0.1
Training code: GitHub

License

Apache 2.0 (inherits from Qwen3 base model).